diff --git a/datasets/oplm-antibody-sequences/SKILL.md b/datasets/oplm-antibody-sequences/SKILL.md index 20c6d75..d0ae198 100644 --- a/datasets/oplm-antibody-sequences/SKILL.md +++ b/datasets/oplm-antibody-sequences/SKILL.md @@ -1,111 +1,112 @@ --- name: oplm-antibody-sequences -description: "Human antibody heavy and light chains from AntiRef v2022.12.14 (OAS, healthy donors' blood), laid out as OPLM's sequence data: AntiRef96 clusters with their AntiRef99 members, AntiRef99 representatives, and held-out AntiRef90 clusters with locus and an IMGT CDR mask. Reach for it to add antibodies to OPLM training or to measure CDR3 loss." +description: "Human antibody heavy and light chains from AntiRef v2022.12.14 (OAS, healthy donors' blood), laid out as OPLM's sequence data, with a heavy-chain CDR3-identity holdout at 85%: 5,000 held-out heavy chains whose CDR3s have no 85%-identical CDR3 left in training, 1.0.0's AntiRef90-cluster eval and test files unchanged, and four screened train sets (AntiRef90, 92, 94 and 96 clusters with AntiRef99 members). Reach for it to compare antibody train sets in OPLM's anneal, or to measure CDR3 loss on CDR3s the model has not seen." metadata: - version: 1.0.0 + version: 2.0.0 tools: datascience@1.2.0 license: cc-by-4.0 --- # oplm-antibody-sequences -Hub: `brineylab/oplm-antibody-sequences`, tag `v1.0.0` +Hub: `brineylab/oplm-antibody-sequences`, tag `v2.0.0` -A pilot: does a decay with antibodies improve OPLM's antibody performance, judged on CDR3, against a decay on protein sequences only? It is built from the published AntiRef as it stands (Briney 2023, *Bioinformatics Advances*, [10.1093/bioadv/vbad109](https://doi.org/10.1093/bioadv/vbad109)). A rebuild from current OAS, with donors and germline annotations, waits on the pilot. +Version 1.0.0 held out whole AntiRef90 clusters and nothing more, so 27% of eval_gte2's heavy chains had an exact copy of their CDR3 in train. Version 2.0.0 adds a holdout where "unseen" means unseen: no heavy-chain CDR3 left in training is 85% or more identical to a held-out heavy chain's CDR3. Its four train sets share that holdout and that screen, so OPLM decays that differ only in train set can be compared (briney/oplm#78). It is built from the published AntiRef (Briney 2023, *Bioinformatics Advances*, [10.1093/bioadv/vbad109](https://doi.org/10.1093/bioadv/vbad109)). A donor holdout waits on a rebuild from current OAS. ## Files -The layout is `brineylab/oplm-sequence-data`'s, one folder per training source: - ``` -antiref96/train/shard-00000.parquet … one row per AntiRef96 cluster -antiref99/train/shard-00000.parquet … one row per AntiRef99 representative +antiref90/train/shard-00000.parquet … one row per AntiRef90 cluster +antiref92/train/ … antiref94/train/ … antiref96/train/ … +eval_cdrh3.parquet test_cdrh3.parquet eval_singletons.parquet eval_gte2.parquet test_singletons.parquet test_gte2.parquet ``` -- **`antiref96/train/`**, uniref70's columns: - - `sequence_id`: the AntiRef96 representative's AntiRef UUID; - - `sequence`: its amino acids; - - `member_ids`, `member_sequences`: the AntiRef99 representatives inside the cluster, the representative's own entry first. -- **`antiref99/train/`**, deepclust70's columns: `sequence_id`, `sequence`. -- **Shards:** 1,000,000 rows each (the last fewer), row groups of 10,000, zstd. Rows are in a seeded global shuffle (`md5('1:')`), so any subset of shards is a uniform sample. -- **Eval and test** each hold 5,000 representatives of held-out AntiRef90 clusters: `_singletons` from clusters of one AntiRef100 sequence, `_gte2` from clusters of two or more. Heavy and light chains are mixed. Columns: +- **`antiref{90,92,94,96}/train/`**, uniref70's columns: + - `sequence_id`, `sequence`: the cluster's representative, its AntiRef UUID and amino acids. **A row's `sequence` is AntiRef's representative unless the screen removed it**; then the cluster's first surviving member, in the seeded order, takes its place; + - `member_ids`, `member_sequences`: the surviving AntiRef99 representatives inside the cluster, the row's own entry first. + + All four folders hold the same 299,748,717 AntiRef99 sequences, grouped differently, so the four train sets differ only in how training draws from them. +- **Shards:** 1,000,000 rows each (the last fewer), row groups of 10,000, zstd. Rows are in a seeded global shuffle (`md5('1:')`), so any subset of shards is a uniform sample. Members follow the representative in `md5('1:')` order. +- **`eval_cdrh3`, `test_cdrh3`:** 2,500 heavy chains each, the CDR3-identity holdout. Every CDR3 has at least 5 residues, and no two CDR3s across both files are 85% identical. +- **`eval_singletons`, `eval_gte2`, `test_singletons`, `test_gte2`:** 1.0.0's files, byte-identical, so earlier #78 numbers stay comparable. 5,000 representatives of held-out AntiRef90 clusters each, heavy and light mixed. +- **Held-out columns**, all six files: - `sequence_id`, `sequence`; - - `locus`: `IGH`, `IGK`, or `IGL`; - - `cdr_mask_aa`: abstar 0.9.0's CDR mask, one digit per residue of `sequence`: `0` framework, `1` CDR1, `2` CDR2, `3` CDR3. It uses IMGT regions (CDR1 27–38, CDR2 56–65, CDR3 105–117, insertions with their base position) from ANARCII 2.0.8's numbering. + - `locus`: `IGH`, `IGK`, or `IGL` (`IGH` only in the `_cdrh3` files); + - `cdr_mask_aa`: one digit per residue of `sequence`: `0` framework, `1` CDR1, `2` CDR2, `3` CDR3, from ANARCII 2.0.8's IMGT numbering (CDR1 27–38, CDR2 56–65, CDR3 105–117). - Every member of a held-out cluster is out of both train sets. + Every member of a held-out AntiRef90 cluster is out of every train set. -`sequence` is the variable domain, FWR1 through FWR4: the span abstar's `sequence_vdj_aa` covers. +`sequence` is the variable domain, FWR1 through FWR4. ## Reading it ```python from huggingface_hub import snapshot_download -path = snapshot_download("brineylab/oplm-antibody-sequences", repo_type="dataset", revision="v1.0.0") +path = snapshot_download("brineylab/oplm-antibody-sequences", repo_type="dataset", revision="v2.0.0") ``` -In OPLM, a `data.train` source is a folder of shards: -- **AntiRef96 representatives:** `antiref96/train` with `sample_members: false`. -- **AntiRef96 clusters sampled to one AntiRef99 member:** `antiref96/train` with `sample_members: true`. -- **AntiRef99 representatives:** `antiref99/train`. - -The four held-out files are `type: sequence` evals. The CDR3 eval, which scores only positions where `cdr_mask_aa` is `3` and runs per `locus`, is briney/oplm#85. +In OPLM, a `data.train` source is one of the four folders, **always with `sample_members: true`**: a row's representative may stand in for a removed one, so the representatives alone are not AntiRef's. The six held-out files are `type: sequence` evals; OPLM's `antibody_cdr3` eval reads them unchanged and scores only positions where `cdr_mask_aa` is `3`, per `locus`. Read heavy-only metrics on the `_cdrh3` files. ## Recipe -**Source:** AntiRef v2022.12.14 on Zenodo (CC BY 4.0), pulled 2026-11-09: +**Source:** AntiRef v2022.12.14 on Zenodo (CC BY 4.0), pulled 2026-11-09, the same two files as 1.0.0: - record 7474336, `antiref_cluster-manifest.csv.gz` (14,054,641,207 bytes, MD5 `18fcf6376702a2d23e2acdc32f1fd40a`): one row per AntiRef100 sequence, with its representative at 100, 99, 98, 96, 94, 92, and 90% identity; - record 7474658, `antiref100.fasta.gz` (18,496,427,276 bytes, MD5 `4a82d4be0c493377668fac9730aca866`): each sequence's UUID and amino acids. The recipe refuses a download whose MD5 differs. AntiRef's clustering was MMseqs2 linclust, nested: each threshold clustered the previous one's representatives, with `-c 0.8 --cov-mode 0`. -**Steps**, one job in `datascience@1.2.0`: +**Identity.** Two CDR3s `a` and `b` are 85% identical when `levenshtein(a, b) <= max(len(a), len(b)) * 3 // 20`. Length differences count against identity. + +**Steps**, one job in `datascience@1.2.0`. Each stage lives under `/.work/` and is skipped when a rerun finds it finished. A self-check of the matching code against known answers runs first. + +1. **Stage** the manifest and the FASTA as parquet, as 1.0.0. +2. **1.0.0's holdout**, unchanged (`seed = 1`, `heldout = 5000`): its four files come out byte-identical. +3. **The heavy CDR3 holdout.** Walk the remaining AntiRef90 representatives in `md5(':')` order and number each with ANARCII (IMGT). Keep a candidate when its chain is H, every residue is numbered, its CDR3 has at least 5 residues, and its CDR3 is not 85% identical to one already kept. The first 2,500 kept are eval, the next 2,500 test. +4. **The screen.** Every AntiRef99 sequence outside the held-out AntiRef90 clusters is checked against the 5,000 held-out CDR3s, on its C-terminal half (`sequence[len(sequence) // 2:]`), which always holds the CDR3: + - **anchored:** every substring of 3–40 residues right after a `C` and right before a `[WF]G.G` is a candidate CDR3. The sequence goes when any candidate is 85% identical to a held-out CDR3. Trying every boundary pair means a misplaced boundary can't hide a match, and matching end to end keeps shared D or J stretches from removing unrelated antibodies; + - **unanchored** (no candidate): the sequence goes when a held-out CDR3 `q` aligns semi-globally anywhere in the half within `len(q) * 3 // 17` edits, a superset of the 85% bound for any substring length. -1. **Stage** the manifest and the FASTA as parquet. Every manifest ID must have a sequence, and every sequence a manifest row. -2. **Hold out.** - - Put the AntiRef90 clusters in the seeded order, `md5(':')`, and walk singletons and gte2 separately. - - Number each representative with ANARCII (`seq_type="antibody"`, `mode="accuracy"`, IMGT). Keep it when its chain type is H, K, or L, every residue is numbered, and it has a CDR3. - - The first `heldout` kept are eval, the next `heldout` test. Candidates that fail numbering stay in train; the build prints how many. - - Before any of this, the build numbers one known heavy chain and stops if its mask moved. -3. **Train.** Drop every manifest row in a held-out cluster. AntiRef99 rows are the distinct `antiref99`. AntiRef96 rows are the distinct `antiref96`, with their distinct `antiref99` as members; the build stops if a representative is not its own first member. Both are written in the seeded order. + A pigeonhole index (an exact piece of the query must appear in any match) narrows the candidates before alignment. Hits are written to `.work/screened.parquet`, then the tripwire is checked: the build stops if the screen removed more than `screen_max` of AntiRef99 sequences. A rerun with a raised `screen_max` reuses the hits. +5. **Train.** Drop held-out clusters and screened sequences, then write each level's clusters with their surviving AntiRef99 members, in the seeded order; a cluster with no survivor is dropped. The build stops if a representative is not among its own AntiRef99 members, or if the four folders' member totals differ. `params.toml`: -- `seed = 1`, `heldout = 5000`, `shard_rows = 1000000`; +- `seed = 1`, `heldout = 5000`: 1.0.0's holdout; +- `cdrh3_seed = 2`, `cdrh3_heldout = 2500`; +- `screen_max = 0.06`; +- `shard_rows = 1000000`; - `limit = 0`: whole AntiRef90 clusters, for a developer's dry run. 0 is all, and the published version has 0. -**The test** checks: -- the files, their columns, and `heldout` rows in each held-out file; +**The test** checks, with its own matching code (DuckDB's `levenshtein` and a q-gram-filtered semi-global search), not the build's: +- the files, their columns and row counts; 1.0.0's four by sha256; - the shards' names and sizes, and each train folder against its floor; -- no `sequence_id` in two held-out files or in a held-out file and a train shard (first and last); -- each mask's length, digits, and single CDR3 run; -- `locus`; -- representative-first members. +- `_cdrh3` rows are `IGH` with a CDR3 of at least 5; masks' length, digits and single CDR3 run; +- no two held-out CDR3s 85% identical; +- no `sequence_id` in two held-out files, or in a held-out file and a train shard (first and last of each folder); +- representative-first members; +- **the screen holds** on 1,000 rows of each folder's first shard: no anchored candidate is 85% identical to a held-out CDR3, and 200 unanchored members carry none by the semi-global search. ## Compute -Measured on the first full build, on a workstation under podman with 16 CPUs and 100 GB of memory: -- **Time:** about 37 minutes. - - Download, 32 parallel range requests: 11 min. - - Stage: 9 min. - - Hold out (ANARCII on 16 CPUs): 5 min. - - AntiRef99: 3 min. - - AntiRef96: 9 min. -- **Memory:** the container's peak reached its 100 GB limit. That peak counts page cache, and DuckDB spills past 80% of the limit, so 100 GB is enough but not much more than enough; ask 128 GB. -- **Disk:** - - `.work/`: 89 GB (32.5 GB of downloads and 62.5 GB staged); - - DuckDB's spill: up to about 90 GB; - - the output: 40 GB (AntiRef96 24 GB, AntiRef99 16 GB). -- **Rows:** AntiRef96 210,559,202 in 211 shards; AntiRef99 315,983,206 in 316 shards. The test's floors are 90% of these. +Measured on the first full build, on bcr's `cpu` partition with 32 CPUs and 256 GB of memory, run as flyloft jobs: +- **Time:** about 7.5 hours from nothing. + - Download, stage and 1.0.0's holdout: about 50 min. + - The CDR3 holdout: 5 min (11,478 candidates tried, 275 rejected as near a kept CDR3). + - The screen: about 5.7 hours. Real data ran about four times slower than the synthetic benchmark (217 µs per anchored sequence): public CDR3s make many full alignments. + - The four train folders and the test: about 45 min. +- **Memory:** not measured (flyloft did not return `sacct`); 256 GB was enough. +- **Disk:** `.work/` 117 GB; the output 80 GB. +- **The screen:** removed 16,222,603 of 315,971,320 AntiRef99 sequences (5.134%). 1,472 held-out CDR3s each removed more than 1,000; the largest removed 358,021. +- **Rows** (mean members per row): AntiRef90 152,351,629 in 153 shards (1.97); AntiRef92 161,836,404 in 162 (1.85); AntiRef94 175,083,828 in 176 (1.71); AntiRef96 199,443,463 in 200 (1.50). The test's floors are 90% of these. ## Gotchas -- **Zenodo is slow to CoreWeave.** On 2026-11-10 it served CoreWeave at 5–7 KB/s per connection and the Scripps network at about 520 KB/s, while CoreWeave reached other hosts at hundreds of MB/s. The 32.5 GB download would take about two days on CoreWeave, so build on bcr. -- **Held out by AntiRef90 cluster membership only.** No identity screen was run. - - A train sequence can share a held-out sequence's V gene and a similar CDR3 at under 90% identity. - - Because linclust is greedy and nested over six levels, a train sequence can even be 90% or more identical to a held-out representative and sit in another AntiRef90 cluster. - - Antibody loss on these sets flatters the model. Compare runs on the same sets; don't read absolute values. -- Two clustering thresholds live in one dataset, against the datasets design's rule (D3), by the lab's decision for this pilot: every train option shares one set of eval and test files. -- No screen against OPLM's uniref70 or deepclust70 either: an antibody from UniProt may sit in both. +- **From 1.0.0:** `antiref99/train` is gone; `antiref96/train` is screened and can promote a member to representative; `antiref90/92/94/train` and the `_cdrh3` files are new; the four AntiRef90-cluster files are unchanged. Train results are not comparable across the two versions. +- **Against the datasets design's rule (D3), by the PI's decisions:** four clustering levels live in one dataset (2026-11-12, extending 1.0.0's exception), since all four train sets must share one holdout and one screen; and the new holdout is a version of this dataset, not a new one (2026-11-13), since it is the same data curated with a different eval and test holdout. +- **The tripwire was raised to 6%** by the PI on 2026-11-13, after the first build removed 5.134% against the planned 5%. The held-out sets were kept as drawn: public CDR3s are the ones whose holdout matters most. There is no per-CDR3 cap, which would bias the eval away from public CDR3s. +- **Light chains are held out by AntiRef90 cluster only**, as in 1.0.0. Their CDR3s are low-diversity; light-chain loss here still rewards recall. +- **The screen is CDR3-only.** A train sequence can share a held-out chain's V gene and framework; only its CDR3 is kept below 85%. +- **The screen over-removes slightly, by design:** every boundary pair is tried, the whole C-terminal half is searched, light chains are screened too, and unanchored sequences get a bound looser than 85%. +- **No screen against OPLM's uniref70 or deepclust70:** an antibody from UniProt may sit in both. - AntiRef's filter: OAS unpaired, human, PBMC, no disease, no vaccine. There is no donor, no nucleotide sequence, no germline call, and no SHM mask; a donor holdout needs a rebuild. -- Light chains are low-diversity and easy for a model; read heavy-only CDR3 loss for the signal. -- In a production decay, 3% of rows exceeds AntiRef99's 316M, so the antibodies repeat. A pilot that subsamples takes a subset of shards. +- **Zenodo is slow to CoreWeave.** On 2026-11-10 it served CoreWeave at 5–7 KB/s per connection and the Scripps network at about 520 KB/s, so build on bcr. +- **A flyloft job's bundle drops the executable bit:** run `python3 -I recipe/build`, not `recipe/build`. diff --git a/datasets/oplm-antibody-sequences/recipe/build b/datasets/oplm-antibody-sequences/recipe/build index cccbde1..50356e4 100755 --- a/datasets/oplm-antibody-sequences/recipe/build +++ b/datasets/oplm-antibody-sequences/recipe/build @@ -1,16 +1,20 @@ #!/usr/bin/env -S python3 -I -"""recipe/build : OPLM's antibody sequences from AntiRef v2022.12.14 (Zenodo 7474336 and 7474658). Held-out -AntiRef90 clusters, singletons and gte2, are numbered with ANARCII and written as eval and test with `locus` and an -abstar-format `cdr_mask_aa`; every other cluster goes to two train sets, AntiRef96 clusters with their AntiRef99 -members and AntiRef99 representatives, sharded as OPLM's uniref70 and deepclust70 are. +"""recipe/build : OPLM's antibody sequences from AntiRef v2022.12.14 (Zenodo 7474336 and 7474658), with a +heavy-chain CDR3-identity holdout. oplm-antibody-sequences v1.0.0's held-out AntiRef90 clusters come out unchanged. Then +10,000 more heavy-chain AntiRef90 representatives, their CDR3s mutually under 85% identical, are eval_cdrh3 and +test_cdrh3, and every AntiRef99 sequence carrying one of those CDR3s at 85% identity or more is screened out of train. +What remains is written as AntiRef90, 92, 94 and 96 clusters with their AntiRef99 members, sharded as OPLM's uniref70. -Stages under /.work/, each skipped when a rerun finds it finished. Prints `held out failed numbering` lines, -then `rows` and `locus` lines last. +Stages under /.work/, each skipped when a rerun finds it finished. Runs a self-check first (`--selfcheck` alone +stops there). Prints `held out failed numbering`, `cdrh3` and `screen` lines, then `rows`, `members` and `locus` lines. """ +import collections import gzip import hashlib import itertools +import multiprocessing import os +import re import shutil import sys import tomllib @@ -30,6 +34,7 @@ SOURCE = {"antiref_cluster-manifest.csv.gz": (7474336, 14054641207, "18fcf637670 STREAMS = 32 LEVELS = ["antiref100", "antiref99", "antiref98", "antiref96", "antiref94", "antiref92", "antiref90"] REGIONS = [(27, 38, "1"), (56, 65, "2"), (105, 117, "3")] # IMGT; everything else is framework, 0 +TRAIN_LEVELS = ["antiref90", "antiref92", "antiref94", "antiref96"] LOCI = {"H": "IGH", "K": "IGK", "L": "IGL"} STRATA = {"singletons": "size = 1", "gte2": "size >= 2"} # a heavy chain whose IMGT regions are FWR1 25, CDR1 8, FWR2 17, CDR2 8, FWR3 38, CDR3 21 (with insertions), FWR4 11 @@ -54,6 +59,139 @@ def connect() -> duckdb.DuckDBPyConnection: return con +def levenshtein(a: str, b: str) -> int: + prev = list(range(len(b) + 1)) + for i, x in enumerate(a, 1): + cur = [i] + for j, y in enumerate(b, 1): + cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (x != y))) + prev = cur + return prev[-1] + + +def within(a: str, b: str, k: int) -> bool: + """levenshtein(a, b) <= k, filling only the band of width 2k + 1 around the diagonal and stopping once a row + exceeds k: the screen's hot path.""" + if abs(len(a) - len(b)) > k: + return False + big = k + 1 + prev = [j if j <= k else big for j in range(len(b) + 1)] + for i, x in enumerate(a, 1): + lo, hi = max(1, i - k), min(len(b), i + k) + cur = [big] * (len(b) + 1) + cur[0] = i if i <= k else big + for j in range(lo, hi + 1): + cur[j] = min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (x != b[j - 1])) + if min(cur[lo - 1:hi + 1]) > k: + return False + prev = cur + return prev[-1] <= k + + +def near(a: str, b: str) -> bool: + """85% CDR3 identity: an edit distance of at most floor(0.15 x the longer length).""" + return within(a, b, max(len(a), len(b)) * 3 // 20) + + +def anchored(sequence: str) -> list[str]: + """Every CDR3 the C-terminal half could hold: 3-40 residues right after a C and right before [WF]G.G.""" + window = sequence[len(sequence) // 2:] + starts = [m.end() for m in re.finditer("C", window)] + ends = [m.start() for m in re.finditer("(?=[WF]G.G)", window)] + return list(dict.fromkeys(window[s:e] for s in starts for e in ends if 3 <= e - s <= 40)) + + +def semiglobal_within(query: str, window: str, k: int) -> bool: + """Sellers: whether some substring of `window` is within k edits of `query`.""" + prev = [0] * (len(window) + 1) + for i, x in enumerate(query, 1): + cur = [i] + for j, y in enumerate(window, 1): + cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (x != y))) + prev = cur + return min(prev) <= k + + +class PieceIndex: + """Pigeonhole index over held-out CDR3s. A query of length L is cut into 3L // 17 + 1 pieces; anything near it + holds one piece exactly, since near allows floor(0.15 x max length) edits and a near string is at most 20L/17 long, + so at most 3L // 17 edits.""" + + def __init__(self) -> None: + self.queries: dict[int, str] = {} + self.pieces: dict[str, set[int]] = {} + self.lengths: set[int] = set() + self.memo: dict[str, int | None] = {} # CDR3s repeat (light chains most), so each is matched once per worker + + def add(self, query: str, qid: int) -> None: + self.memo.clear() + self.queries[qid] = query + p = len(query) * 3 // 17 + 1 + bounds = [round(i * len(query) / p) for i in range(p + 1)] + for lo, hi in itertools.pairwise(bounds): + self.pieces.setdefault(query[lo:hi], set()).add(qid) + self.lengths.add(hi - lo) + + def candidates(self, text: str) -> set[int]: + hits: set[int] = set() + for n in self.lengths: + for i in range(len(text) - n + 1): + hits |= self.pieces.get(text[i:i + n], set()) + return hits + + def match(self, cand: str) -> int | None: + if cand not in self.memo: + if len(self.memo) > 5_000_000: # ponytail: wholesale reset bounds memory; an LRU if hit rates suffer + self.memo.clear() + self.memo[cand] = min((q for q in self.candidates(cand) if near(self.queries[q], cand)), default=None) + return self.memo[cand] + + +def screened_by(idx: PieceIndex, sequence: str) -> int | None: + """The held-out CDR3 (qid) a train sequence carries at 85% identity, or None. Anchored candidates are compared + whole; a sequence with none falls back to a semi-global search of its window within 3 x query length // 17.""" + cands = anchored(sequence) + if cands: + return min((q for c in cands if (q := idx.match(c)) is not None), default=None) + window = sequence[len(sequence) // 2:] + return next((q for q in sorted(idx.candidates(window)) + if semiglobal_within(idx.queries[q], window, len(idx.queries[q]) * 3 // 17)), None) + + +def selfcheck() -> None: + """Known answers for the CDR3 matching, run before every build. KNOWN's CDR3 is 21 residues, so 3 edits are near.""" + q = "ARSPYSSSWPGYYYYYYGMDV" + assert levenshtein("kitten", "sitting") == 3 + for x, y in [("kitten", "sitting"), ("", "abc"), ("abc", ""), (q, q[3:]), (q, "GRSPYSASWPGYYAYYYGMDI"), + ("ARDYYYYYGMDV", "ARGGGSDYYYYYGMDVTT"), ("ACDEFG", "GFEDCA"), ("AAAA", "AAAAA")]: + for k in range(0, 8): + assert within(x, y, k) == (levenshtein(x, y) <= k), (x, y, k) + assert near(q, q) and near(q, q[:-1] + "I") + assert near(q, "GRSPYSASWPGYYAYYYGMDV") # 3 substitutions + assert not near(q, "GRSPYSASWPGYYAYYYGMDI") # 4 substitutions: one past the bound + assert not near(q, "GRSQYSASWQGYYAYYYGMDI") # 6 substitutions + assert not near("ARDYYYYYGMDV", "ARGGGSDYYYYYGMDVTT") # 6 apart, the bound is 3 + assert near("", "") and anchored("") == [] and anchored("nan") == [] + assert q in anchored(KNOWN[0]) + w = KNOWN[0][len(KNOWN[0]) // 2:] + assert semiglobal_within(q, w, 5) and not semiglobal_within("WWWWWWWWWW", w, 2) + idx = PieceIndex() + idx.add(q, 7) + assert idx.match(q) == 7 and idx.match("GRSPYSASWPGYYAYYYGMDV") == 7 + assert idx.match("ARDYYYYYGMDV") is None + loose = KNOWN[0].replace("WGQG", "WAQA") # no [WF]G.G anchor, so only the fallback can find it + assert screened_by(idx, KNOWN[0]) == 7 + assert anchored(loose) == [] and screened_by(idx, loose) == 7 + assert screened_by(idx, KNOWN[0].replace(q, "AKDLGGNWFDP")) is None + assert promote("antiref90", "A", "A", False) == "A" + assert promote("antiref90", "A", "B", True) == "B" # the screen took the representative: its first survivor leads + try: + promote("antiref90", "A", "B", False) + raise AssertionError("a broken nesting passed") + except SystemExit: + pass + + def fetch_range(url: str, part: Path, lo: int, hi: int) -> None: request = urllib.request.Request(url, headers={"Range": f"bytes={lo}-{hi}"}) with urllib.request.urlopen(request) as r, part.open("r+b") as f: @@ -168,6 +306,78 @@ def heldout_sets(work: Path, n: int) -> dict[str, list[dict]]: return sets +def cdrh3_heldout(work: Path, seed: int, n: int, exclude: set[str]) -> dict[str, list[dict]]: + """Heavy-chain AntiRef90 representatives outside `exclude`, in the seeded order md5(':'), kept + when numbered with a CDR3 of 5 or more residues that is under 85% identical to every CDR3 kept before it. The + first n kept are eval_cdrh3, the next n test_cdrh3.""" + anarcii = Anarcii(seq_type="antibody", mode="accuracy", cpu=True, ncpu=cpus) + rows = connect().execute( + f"SELECT m.antiref90, s.sequence FROM (SELECT DISTINCT antiref90 FROM '{work / 'manifest.parquet'}') m " + f"JOIN '{work / 'sequences.parquet'}' s ON s.sequence_id = m.antiref90 ORDER BY md5('{seed}:' || m.antiref90)") + idx, kept, tried, rejected = PieceIndex(), [], 0, 0 + while len(kept) < 2 * n and (pool := rows.fetchmany(int(2.4 * n))): + pool = [(sid, seq) for sid, seq in pool if sid not in exclude] + numbered = anarcii.number(dict(pool)) + for sid, sequence in pool: + if len(kept) == 2 * n: + break + tried += 1 + result = numbered.get(sid, {"chain_type": "F"}) + mask = cdr_mask(result["numbering"], sequence) if result["chain_type"] == "H" else None + cdr3 = "".join(a for a, d in zip(sequence, mask or "") if d == "3") + if len(cdr3) < 5: + continue + if idx.match(cdr3) is not None: + rejected += 1 + continue + idx.add(cdr3, len(kept)) + kept.append({"sequence_id": sid, "sequence": sequence, "locus": "IGH", "cdr_mask_aa": mask}) + if len(kept) < 2 * n: + sys.exit(f"cdrh3: {len(kept)} of {2 * n} after {tried} candidates") + print(f"cdrh3 candidates tried: {tried}") + print(f"cdrh3 rejected near: {rejected}") + return {"eval_cdrh3": kept[:n], "test_cdrh3": kept[n:]} + + +_screen_idx: PieceIndex | None = None + + +def _screen_group(task: tuple[str, int]) -> list[tuple[str, int]]: + path, group = task + t = pq.ParquetFile(path).read_row_group(group, use_threads=False) + return [(sid, q) for sid, seq in zip(t["antiref99"].to_pylist(), t["sequence"].to_pylist()) + if (q := screened_by(_screen_idx, seq)) is not None] + + +def screen(work: Path, queries: list[str]) -> None: + """Every AntiRef99 sequence outside the held-out AntiRef90 clusters that carries a held-out heavy CDR3 at 85% + identity, as .work/screened.parquet (antiref99, qid). Written before the tripwire is checked, so a rerun with a + raised `screen_max` reuses it.""" + global _screen_idx + src = work / "screen_in.parquet" + if not src.is_file(): + con = connect() + con.execute(f"COPY (SELECT d.antiref99, s.sequence FROM (SELECT DISTINCT antiref99 FROM '{work / 'manifest.parquet'}' " + f"ANTI JOIN '{work / 'heldout.parquet'}' USING (antiref90)) d JOIN '{work / 'sequences.parquet'}' s " + f"ON s.sequence_id = d.antiref99) TO '{src}.part' (FORMAT parquet, ROW_GROUP_SIZE 100000)") + con.close() + os.rename(f"{src}.part", src) + _screen_idx = PieceIndex() + for qid, q in enumerate(queries): + _screen_idx.add(q, qid) + meta = pq.ParquetFile(src).metadata + # fork, not spawn: spawn would rerun this script's top level in every worker + with multiprocessing.get_context("fork").Pool(cpus) as pool: + hits = [h for part in pool.imap_unordered(_screen_group, [(str(src), g) for g in range(meta.num_row_groups)]) + for h in part] + total, per_query = meta.num_rows, collections.Counter(q for _, q in hits) + print(f"screen removed: {len(hits)} of {total} ({100 * len(hits) / total:.3f}%)") + print(f"screen cdr3s over 1000: {sum(c > 1000 for c in per_query.values())}") + print(f"screen largest: {max(per_query.values(), default=0)}") + pq.write_table(pa.table({"antiref99": [h[0] for h in hits], "qid": pa.array([h[1] for h in hits], pa.int32())}, + schema=pa.schema([("antiref99", pa.string()), ("qid", pa.int32())])), work / "screened.parquet") + + def write_shards(batches: pa.RecordBatchReader, folder: Path, shard_rows: int) -> None: tmp = folder.with_suffix(".part") shutil.rmtree(tmp, ignore_errors=True) @@ -188,63 +398,87 @@ def write_shards(batches: pa.RecordBatchReader, folder: Path, shard_rows: int) - tmp.rename(folder) -def clusters(reader: pa.RecordBatchReader): - """One row per AntiRef96 cluster from rows (antiref96, antiref99, sequence) sorted so that each cluster's rows are - together, its representative first. A batch's last cluster may go on in the next batch, so it waits for it.""" +def promote(level: str, level_id: str, first_member: str, rep_removed: bool) -> str: + """A cluster row's sequence_id: its representative, or, when the screen removed that, its first surviving member. + A representative that survived but is not its own first member means the nesting broke, and the build stops.""" + if first_member != level_id and not rep_removed: + sys.exit(f"{level} representative {level_id} is not among its own AntiRef99 members") + return first_member + + +def clusters(reader: pa.RecordBatchReader, level: str): + """One row per `level` cluster from rows (level, antiref99, sequence, rep_removed) sorted so that each cluster's + rows are together, its representative first when it survived. A batch's last cluster may go on in the next batch, + so it waits for it.""" held = reader.schema.empty_table() for batch in itertools.chain(reader, [None]): t = held if batch is None else pa.concat_tables([held, pa.Table.from_batches([batch])]).combine_chunks() if not t.num_rows: continue - c = t["antiref96"].combine_chunks() + c = t[level].combine_chunks() starts = [0, *pc.add(pc.indices_nonzero(pc.not_equal(c[1:], c[:-1])), 1).to_pylist()] end = t.num_rows if batch is None else starts.pop() held, t = t.slice(end), t.slice(0, end) if not starts: continue ids, seqs, offsets = t["antiref99"].combine_chunks(), t["sequence"].combine_chunks(), pa.array([*starts, end], pa.int32()) - stray = pc.not_equal(c.take(starts), ids.take(starts)) - if pc.any(stray).as_py(): - sys.exit(f"AntiRef96 representative {c.take(starts).filter(stray)[0]} is not among its own AntiRef99 members") - yield pa.record_batch({"sequence_id": c.take(starts), "sequence": seqs.take(starts), + reps, firsts, removed = c.take(starts), ids.take(starts), t["rep_removed"].combine_chunks().take(starts) + for i in pc.indices_nonzero(pc.not_equal(reps, firsts)).to_pylist(): + promote(level, reps[i].as_py(), firsts[i].as_py(), removed[i].as_py()) + yield pa.record_batch({"sequence_id": firsts, "sequence": seqs.take(starts), "member_ids": pa.ListArray.from_arrays(offsets, ids), "member_sequences": pa.ListArray.from_arrays(offsets, seqs)}) def train(work: Path, shard_rows: int) -> None: - """Every manifest row outside the held-out AntiRef90 clusters, as AntiRef96 clusters with their AntiRef99 members - (representative first) and as AntiRef99 representatives, each in the seeded order.""" + """The AntiRef99 sequences outside the held-out AntiRef90 clusters and outside the screen, as clusters at each of + TRAIN_LEVELS with their AntiRef99 members (representative, or its first survivor, first), in the seeded order.""" con = connect() con.execute(f"CREATE VIEW kept AS SELECT * FROM '{work / 'manifest.parquet'}' ANTI JOIN '{work / 'heldout.parquet'}' h " - f"USING (antiref90)") + f"USING (antiref90) ANTI JOIN '{work / 'screened.parquet'}' r USING (antiref99)") con.execute(f"CREATE VIEW seqs AS SELECT * FROM '{work / 'sequences.parquet'}'") - if not (out / "antiref99" / "train").is_dir(): - write_shards(con.sql(f"SELECT d.antiref99 AS sequence_id, s.sequence FROM (SELECT DISTINCT antiref99 FROM kept) d " - f"JOIN seqs s ON s.sequence_id = d.antiref99 ORDER BY {order('d.antiref99')}").to_arrow_reader(), - out / "antiref99" / "train", shard_rows) - if not (out / "antiref96" / "train").is_dir(): + members = {} + for level in TRAIN_LEVELS: + members[level] = con.sql(f"SELECT count(*) FROM (SELECT DISTINCT {level}, antiref99 FROM kept)").fetchall()[0][0] + if (out / level / "train").is_dir(): + continue # one sort and a streaming group: an ordered list() over 300M sequences spilled 90 GB and stalled - rows = con.sql(f"SELECT m.antiref96, m.antiref99, s.sequence FROM (SELECT DISTINCT antiref96, antiref99 FROM kept) m " - f"JOIN seqs s ON s.sequence_id = m.antiref99 " - f"ORDER BY {order('m.antiref96')}, m.antiref96, m.antiref99 = m.antiref96 DESC, {order('m.antiref99')}") - write_shards(clusters(rows.to_arrow_reader()), out / "antiref96" / "train", shard_rows) - + rows = con.sql(f"SELECT m.{level}, m.antiref99, s.sequence, r.antiref99 IS NOT NULL AS rep_removed " + f"FROM (SELECT DISTINCT {level}, antiref99 FROM kept) m JOIN seqs s ON s.sequence_id = m.antiref99 " + f"LEFT JOIN '{work / 'screened.parquet'}' r ON r.antiref99 = m.{level} " + f"ORDER BY {order(f'm.{level}')}, m.{level}, m.antiref99 = m.{level} DESC, {order('m.antiref99')}") + write_shards(clusters(rows.to_arrow_reader(), level), out / level / "train", shard_rows) + if len(set(members.values())) != 1: + sys.exit(f"member totals differ across levels: {members}") + for level in TRAIN_LEVELS: + rows = con.sql(f"SELECT count(*) FROM '{out / level / 'train'}/*.parquet'").fetchall()[0][0] + print(f"rows {level}/train: {rows}") + print(f"members {level}/train: {members[level]} mean {members[level] / rows:.2f}") +selfcheck() +if sys.argv[1:] == ["--selfcheck"]: + sys.exit(0) download(work) stage(work, params["limit"]) if not (work / "heldout.parquet").is_file(): for name, rows in heldout_sets(work, params["heldout"]).items(): pq.write_table(pa.Table.from_pylist(rows), out / f"{name}.parquet", compression="zstd") ids = [r["sequence_id"] for name in ("eval", "test") for s in STRATA for r in pq.read_table(out / f"{name}_{s}.parquet").to_pylist()] + for name, rows in cdrh3_heldout(work, params["cdrh3_seed"], params["cdrh3_heldout"], set(ids)).items(): + pq.write_table(pa.Table.from_pylist(rows), out / f"{name}.parquet", compression="zstd") + ids += [r["sequence_id"] for r in rows] pq.write_table(pa.table({"antiref90": ids}), work / "heldout.parquet") +if not (work / "screened.parquet").is_file(): + held = [r for name in ("eval_cdrh3", "test_cdrh3") for r in pq.read_table(out / f"{name}.parquet").to_pylist()] + screen(work, ["".join(a for a, d in zip(r["sequence"], r["cdr_mask_aa"]) if d == "3") for r in held]) +removed = pq.ParquetFile(work / "screened.parquet").metadata.num_rows / pq.ParquetFile(work / "screen_in.parquet").metadata.num_rows +if removed > params["screen_max"]: + sys.exit(f"screen removed {100 * removed:.3f}% > {100 * params['screen_max']:g}%") train(work, params["shard_rows"]) con = connect() -for folder in ("antiref96", "antiref99"): - print(f"rows {folder}/train: {con.sql(f"SELECT count(*) FROM '{out / folder / 'train'}/*.parquet'").fetchall()[0][0]}") -for name in ("eval", "test"): - for s in STRATA: - path = out / f"{name}_{s}.parquet" - print(f"rows {path.name}: {pq.ParquetFile(path).metadata.num_rows}") - for locus, count in con.sql(f"SELECT locus, count(*) FROM '{path}' GROUP BY locus ORDER BY locus").fetchall(): - print(f"locus {path.name} {locus}: {count}") +for name in [f"{n}_{st}" for n in ("eval", "test") for st in STRATA] + ["eval_cdrh3", "test_cdrh3"]: + path = out / f"{name}.parquet" + print(f"rows {path.name}: {pq.ParquetFile(path).metadata.num_rows}") + for locus, count in con.sql(f"SELECT locus, count(*) FROM '{path}' GROUP BY locus ORDER BY locus").fetchall(): + print(f"locus {path.name} {locus}: {count}") diff --git a/datasets/oplm-antibody-sequences/recipe/params.toml b/datasets/oplm-antibody-sequences/recipe/params.toml index ce9b566..f2921b7 100644 --- a/datasets/oplm-antibody-sequences/recipe/params.toml +++ b/datasets/oplm-antibody-sequences/recipe/params.toml @@ -1,5 +1,8 @@ pulled = "2026-11-09" -seed = 1 +seed = 1 # v1's holdout, unchanged heldout = 5000 +cdrh3_seed = 2 +cdrh3_heldout = 2500 +screen_max = 0.06 # the build stops if the screen removes more than this share of AntiRef99 train sequences shard_rows = 1000000 limit = 0 # AntiRef90 clusters, whole, for a developer's dry run; 0 is all, and the published version has 0 diff --git a/datasets/oplm-antibody-sequences/recipe/test b/datasets/oplm-antibody-sequences/recipe/test index 5f797f7..db983e0 100755 --- a/datasets/oplm-antibody-sequences/recipe/test +++ b/datasets/oplm-antibody-sequences/recipe/test @@ -1,24 +1,63 @@ #!/usr/bin/env -S python3 -I -"""recipe/test : exit 0 iff holds a version of oplm-antibody-sequences. Reads parquet footers, the four -held-out files, and the first and last shard of each train folder; one line per failure, naming the check and the -first offending sequence_id.""" +"""recipe/test : exit 0 iff holds a version of oplm-antibody-sequences from 2.0.0 on. Reads parquet footers, the six held-out +files, and the first and last shard of each train folder; checks the CDR3 holdout's 85% separation and, on sampled +train members, that the screen holds, with its own matching code (DuckDB's levenshtein and a q-gram-filtered +semi-global search), not the build's. One line per failure, naming the check and the first offender.""" +import hashlib +import re import sys import tomllib from pathlib import Path +import duckdb import polars as pl import pyarrow.parquet as pq params = tomllib.loads((Path(__file__).parent / "params.toml").read_text()) -FLOOR = {"antiref96": 189_503_281, "antiref99": 284_384_885} # rows: 90% of the first full build's -HELDOUT = ["eval_singletons", "eval_gte2", "test_singletons", "test_gte2"] -TRAIN_COLUMNS = {"antiref96": ["sequence_id", "sequence", "member_ids", "member_sequences"], - "antiref99": ["sequence_id", "sequence"]} +TRAIN_LEVELS = ["antiref90", "antiref92", "antiref94", "antiref96"] +FLOOR = {"antiref90": 137116466, "antiref92": 145652763, "antiref94": 157575445, "antiref96": 179499116} # rows: 90% of the first full build's +V1 = {"eval_singletons": "f28027d3b65fc11e1ac790674d5be1885775973c072da9dfd4e4e91899648552", + "eval_gte2": "862f73fe8c1faadcf98c50c5b6c83742dac6e6fc36c1b1331ee4896e6ee28c99", + "test_singletons": "cc6381e9e0fedb8c447925c9d369ef57e219237ab61f60055166b27ed8688758", + "test_gte2": "9456f11b468f016492fc58f03817de55a8b385879a73ded08ee50a92073aa9d5"} # oplm-antibody-sequences v1.0.0 +CDRH3 = ["eval_cdrh3", "test_cdrh3"] +TRAIN_COLUMNS = ["sequence_id", "sequence", "member_ids", "member_sequences"] HELDOUT_COLUMNS = ["sequence_id", "sequence", "locus", "cdr_mask_aa"] +SAMPLE_ROWS, SAMPLE_UNANCHORED = 1000, 200 +NEAR = ("abs(length(a.cdr3) - length(b.cdr3)) <= greatest(length(a.cdr3), length(b.cdr3)) * 3 // 20 " + "AND levenshtein(a.cdr3, b.cdr3) <= greatest(length(a.cdr3), length(b.cdr3)) * 3 // 20") # 85% CDR3 identity +published = not params["limit"] and params["heldout"] == 5000 out, problems = Path(sys.argv[1]), [] +con = duckdb.connect() + + +def cdr3(sequence: str, mask: str) -> str: + return "".join(a for a, d in zip(sequence, mask) if d == "3") + + +def anchored(sequence: str) -> set[str]: + window = sequence[len(sequence) // 2:] + return {window[c.end():w.start()] for c in re.finditer("C", window) for w in re.finditer("(?=[WF]G.G)", window) + if 3 <= w.start() - c.end() <= 40} + + +def sellers(query: str, window: str) -> int: + prev = [0] * (len(window) + 1) + for i, x in enumerate(query, 1): + cur = [i] + for j, y in enumerate(window, 1): + cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (x != y))) + prev = cur + return min(prev) + + +def grams(s: str) -> set[str]: + return {s[i:i + 3] for i in range(len(s) - 2)} + + held = {} -for name in HELDOUT: +for name in [*V1, *CDRH3]: path = out / f"{name}.parquet" if not path.is_file(): problems.append(f"files: {name}.parquet is missing") @@ -27,56 +66,92 @@ for name in HELDOUT: if df.columns != HELDOUT_COLUMNS: problems.append(f"columns: {name}.parquet has {df.columns}, not {HELDOUT_COLUMNS}") continue - if df.height != params["heldout"]: - problems.append(f"rows: {name}.parquet has {df.height}, not {params['heldout']}") + want = params["cdrh3_heldout"] if name in CDRH3 else params["heldout"] + if df.height != want: + problems.append(f"rows: {name}.parquet has {df.height}, not {want}") bad = { - "locus": ~pl.col("locus").is_in(["IGH", "IGK", "IGL"]), + "locus": ~pl.col("locus").is_in(["IGH"] if name in CDRH3 else ["IGH", "IGK", "IGL"]), "cdr_mask_aa length": pl.col("cdr_mask_aa").str.len_chars() != pl.col("sequence").str.len_chars(), "cdr_mask_aa digits": ~pl.col("cdr_mask_aa").str.contains(r"^[0-3]+$"), "cdr_mask_aa one CDR3": ~pl.col("cdr_mask_aa").str.contains(r"^[0-2]*3+[0-2]*$"), } + if name in CDRH3: + bad["CDR3 under 5"] = pl.col("cdr_mask_aa").str.count_matches("3") < 5 for check, expr in bad.items(): hit = df.filter(expr.fill_null(True)) if hit.height: problems.append(f"{check}: {hit.height} rows in {name}.parquet, first {hit['sequence_id'][0]}") - held[name] = set(df["sequence_id"]) + if name in V1 and published and hashlib.sha256(path.read_bytes()).hexdigest() != V1[name]: + problems.append(f"sha256: {name}.parquet differs from oplm-antibody-sequences v1.0.0's") + held[name] = df print(f"{name}.parquet {df.height} rows") +if not published: + print("skip sha256 (a dry run's parameters)") -edges = {} -for folder, columns in TRAIN_COLUMNS.items(): - shards = sorted((out / folder / "train").glob("*.parquet")) +queries = [cdr3(s, m) for name in CDRH3 if name in held for s, m in held[name].select("sequence", "cdr_mask_aa").iter_rows()] +con.execute("CREATE TABLE q AS SELECT unnest(range(?)) AS qid, unnest(?::VARCHAR[]) AS cdr3", [len(queries), queries]) +pair = con.sql(f"SELECT a.qid, b.qid FROM q a JOIN q b ON a.qid < b.qid AND {NEAR} LIMIT 1").fetchall() +if pair: + problems.append(f"cdrh3 separation: held-out CDR3s {queries[pair[0][0]]} and {queries[pair[0][1]]} are 85% identical") + +held_ids = {name: set(df["sequence_id"]) for name, df in held.items()} +members: dict[str, set[str]] = {} +sample: list[str] = [] +for level in TRAIN_LEVELS: + shards = sorted((out / level / "train").glob("*.parquet")) if not shards: - problems.append(f"files: {folder}/train has no shards") + problems.append(f"files: {level}/train has no shards") continue - names = [p.name for p in shards] - if names != [f"shard-{i:05d}.parquet" for i in range(len(shards))]: - problems.append(f"files: {folder}/train's shards are not shard-00000.parquet onward without gaps") + if [p.name for p in shards] != [f"shard-{i:05d}.parquet" for i in range(len(shards))]: + problems.append(f"files: {level}/train's shards are not shard-00000.parquet onward without gaps") counts = [pq.ParquetFile(p).metadata.num_rows for p in shards] if any(n != params["shard_rows"] for n in counts[:-1]): - problems.append(f"rows: a {folder}/train shard before the last is not {params['shard_rows']} rows") - if sum(counts) < FLOOR[folder]: - problems.append(f"rows: {folder}/train has {sum(counts)}, under its floor {FLOOR[folder]}") - for p in {shards[0], shards[-1]}: + problems.append(f"rows: a {level}/train shard before the last is not {params['shard_rows']} rows") + if sum(counts) < FLOOR[level]: + problems.append(f"rows: {level}/train has {sum(counts)}, under its floor {FLOOR[level]}") + for p in sorted({shards[0], shards[-1]}): df = pl.read_parquet(p) - if df.columns != columns: - problems.append(f"columns: {folder}/train/{p.name} has {df.columns}, not {columns}") + if df.columns != TRAIN_COLUMNS: + problems.append(f"columns: {level}/train/{p.name} has {df.columns}, not {TRAIN_COLUMNS}") continue - edges[f"{folder}/train/{p.name}"] = set(df["sequence_id"]) - if folder == "antiref96": - bad = df.filter( - (pl.col("member_ids").list.first() != pl.col("sequence_id")) - | (pl.col("member_sequences").list.first() != pl.col("sequence")) - | (pl.col("member_ids").list.len() != pl.col("member_sequences").list.len()) - ) - if bad.height: - problems.append(f"members: {bad.height} rows in {folder}/train/{p.name}, first {bad['sequence_id'][0]}") - print(f"{folder}/train {len(shards)} shards {sum(counts)} rows") - -seen = {} -for where, ids in [*held.items(), *edges.items()]: - for name, other in seen.items(): - if not (where in edges and name in edges) and ids & other: - problems.append(f"sequence_id: {next(iter(ids & other))} is in {name} and {where}") - seen[where] = ids + bad = df.filter((pl.col("member_ids").list.first() != pl.col("sequence_id")) + | (pl.col("member_sequences").list.first() != pl.col("sequence")) + | (pl.col("member_ids").list.len() != pl.col("member_sequences").list.len()) + | (pl.col("member_ids").list.len() == 0)) + if bad.height: + problems.append(f"members: {bad.height} rows in {level}/train/{p.name}, first {bad['sequence_id'][0]}") + members[f"{level}/train/{p.name}"] = set(df["member_ids"].explode()) + if p == shards[0]: + sample += df.head(SAMPLE_ROWS)["member_sequences"].explode().to_list() + print(f"{level}/train {len(shards)} shards {sum(counts)} rows") + +names = list(held_ids) +for i, a in enumerate(names): + for b in names[i + 1:]: + if held_ids[a] & held_ids[b]: + problems.append(f"sequence_id: {next(iter(held_ids[a] & held_ids[b]))} is in {a} and {b}") + for where, other in members.items(): + if held_ids[a] & other: + problems.append(f"sequence_id: {next(iter(held_ids[a] & other))} is in {a} and {where}") + +# The screen, on sampled members: anchored candidates against every held-out CDR3 in DuckDB; a member with no anchored +# candidate by a semi-global search within 3 x query length // 17, after a q-gram filter (a substring within k +# edits of q keeps all but at most 3k of q's distinct 3-grams: an edit touches three). +cands = [(i, c) for i, s in enumerate(sample) for c in anchored(s)] +con.execute("CREATE TABLE a AS SELECT unnest(?::INTEGER[]) AS i, unnest(?::VARCHAR[]) AS cdr3", + [[i for i, _ in cands], [c for _, c in cands]]) +leak = con.sql(f"SELECT a.i, b.qid FROM a JOIN q b ON {NEAR} LIMIT 1").fetchall() +if leak: + problems.append(f"screen: a sampled member carries held-out CDR3 {queries[leak[0][1]]}: {sample[leak[0][0]]}") +loose = [s for s in sample if not anchored(s)][:SAMPLE_UNANCHORED] +qgrams = [(q, grams(q), len(q) * 3 // 17) for q in queries] +for s in loose: + window = s[len(s) // 2:] + wg = grams(window) + hit = next((q for q, g, k in qgrams if len(g & wg) >= len(g) - 3 * k and sellers(q, window) <= k), None) + if hit: + problems.append(f"screen: unanchored sampled member carries held-out CDR3 {hit}: {s}") + break +print(f"screen sample: {len(sample)} members, {len(cands)} anchored candidates, {len(loose)} unanchored checked") print("\n".join(problems) or "ok") sys.exit(1 if problems else 0) diff --git a/docs/plans/oplm-antibody-cdrh3.md b/docs/plans/oplm-antibody-cdrh3.md new file mode 100644 index 0000000..c574311 --- /dev/null +++ b/docs/plans/oplm-antibody-cdrh3.md @@ -0,0 +1,348 @@ +# `oplm-antibody-cdrh3` Implementation Plan + +> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking. + +**Goal:** Publish `brineylab/oplm-antibody-cdrh3` v1.0.0. It holds a heavy-chain CDR3-identity holdout at 85%, v1's held-out files unchanged, and four screened train sets (AntiRef90, 92, 94 and 96 clusters with AntiRef99 members). Then run briney/oplm#78's four decay branches on it and report. + +**Architecture:** Two repositories, then the cluster. +1. **brineylab/datasets.** One recipe, `datasets/oplm-antibody-cdrh3/recipe/build`, copied from `oplm-antibody-sequences` and extended. It is a single `python3 -I` program in `datascience@1.2.0`, cooked by hand on bcr as v1 was. +2. **CoreWeave.** The sweep reuses `~/projects/oplm/mup_redo/overfit.sh`, extended with `AB_SET=antiref90|92|94|96`. Nothing in briney/oplm's source changes. + +**Tech Stack:** Python 3.12, DuckDB, pyarrow, polars, ANARCII 2.0.8 (all in `datascience@1.2.0`, nothing new); OPLM `antibody_cdr3` eval at main ≥ f28661e; Slurm on bcr and CoreWeave; W&B project `oplm-170M-overfit`. + +**Spec:** `docs/superpowers/specs/2026-11-12-oplm-antibody-cdrh3-design.md` in this worktree (branch `oplm-antibody-cdrh3-design`). Read it first; §5 (steps), §6 (test) and §8 (OPLM) are binding. Also read v1's `datasets/oplm-antibody-sequences/{SKILL.md,recipe/build,recipe/test}`: the new recipe extends them. + +## Global Constraints + +**Revised 2026-11-12 (PI):** 85% CDR3 identity at any length, 2,500 + 2,500 held out, tripwire 5%, raised to 6% on 2026-11-13 after the first full build removed 5.134% (PI). The spec's §2 says why. Tasks 1–4 were first written at 80%; their code is updated to these values. + +**Revised 2026-11-13 (PI), after the first push:** the dataset is published as **`oplm-antibody-sequences` 2.0.0**, not as a new `oplm-antibody-cdrh3`; the recipe and skill live in `datasets/oplm-antibody-sequences/`, the record is `results/oplm-antibody-sequences/2.0.0.txt`, and 2.0.0 has no `antiref99/train`. Wherever a task below says `oplm-antibody-cdrh3` for the Hub repository, the directory or the download path, read `oplm-antibody-sequences` 2.0.0. The candidate pushed to `brineylab/oplm-antibody-cdrh3` (Hub commit 9ab426cb) is abandoned; the PI deletes that repository. + +**Datasets repository** (its `AGENTS.md`, binding): +- Bytes on the Hub, never in git. A recipe never pip-installs. +- Variants are datasets: copy the sibling's directory, change what differs. +- The skill's frontmatter is `name`, `description`, `metadata.version: 1.0.0`, `metadata.tools: datascience@1.2.0`, `metadata.license: cc-by-4.0`. +- The test judges the real output, from footers and samples. +- A commit message is one sentence stating the behavior, ending with the session's attribution line. +- Merge with `gh pr merge --merge` after review by someone other than the builder; then `scripts/promote`. +- Ask the human before every push, PR, Hub upload or promote: these are outward-facing. + +**Source:** v1's two Zenodo files with v1's sizes and MD5s (in v1's `SOURCE`). Build on bcr: Zenodo is slow to CoreWeave. + +**Parameters** (`params.toml`): +```toml +pulled = "2026-11-09" +seed = 1 # v1's holdout, unchanged +heldout = 5000 +cdrh3_seed = 2 +cdrh3_heldout = 2500 +screen_max = 0.06 # build stops if the screen removes more than this share of AntiRef99 train sequences +shard_rows = 1000000 +limit = 0 +``` + +**Definitions** (spec §5.2–5.3, exact): +- `near(a, b)`: `levenshtein(a, b) <= max(len(a), len(b)) * 3 // 20` (that is, `floor(0.15 * max)`). +- Window: `sequence[len(sequence) // 2:]`. +- Anchored candidates: every window substring of 3–40 residues immediately preceded by `C` and immediately followed by a match of `[WF]G.G`. +- A train sequence is screened out if either: + - one of its anchored candidates is `near` a held-out CDR3; or + - it has no anchored candidate, and the semi-global edit distance of a held-out CDR3 `q` against its window is at most `len(q) * 3 // 17`. +- Eval CDR3: residues whose `cdr_mask_aa` digit is `3`. + +**v1 files, byte-identical** (sha256 of v1.0.0): +- `eval_singletons.parquet`: `f28027d3b65fc11e1ac790674d5be1885775973c072da9dfd4e4e91899648552` +- `eval_gte2.parquet`: `862f73fe8c1faadcf98c50c5b6c83742dac6e6fc36c1b1331ee4896e6ee28c99` +- `test_singletons.parquet`: `cc6381e9e0fedb8c447925c9d369ef57e219237ab61f60055166b27ed8688758` +- `test_gte2.parquet`: `9456f11b468f016492fc58f03817de55a8b385879a73ded08ee50a92073aa9d5` + +**Output:** +- **Held-out files.** v1's four, plus `eval_cdrh3.parquet` and `test_cdrh3.parquet`, each with v1's columns (`sequence_id`, `sequence`, `locus`, `cdr_mask_aa`). +- **Train folders.** `antiref{90,92,94,96}/train/shard-%05d.parquet`, each with columns `sequence_id`, `sequence`, `member_ids`, `member_sequences`: + - 1,000,000-row shards, row groups of 10,000, zstd; + - clusters in the order `md5('1:' || )`; + - members ordered representative first, then by `md5('1:' || antiref99)`; + - when the screen removed the representative, the first surviving member leads the row and supplies `sequence_id` and `sequence`. + +**OPLM side:** +- Runs use `overfit.sh decay 140000 25` with `AB=1.5`. +- Run with `PATH=~/venvs/oplm-f28661e/bin:$PATH`. +- Four nodes per run. +- `admin_override` only after the PI approves it at launch. Memory `cluster-qos-override.md` covers small jobs only. + +## Review Focus + +1. **A short or `nan`-bearing CDR3.** + - The ANARCII CDR3 of a candidate can be under 5 residues; it must be rejected, not held out. + - Train sequences contain literal lowercase `nan`. The screen must treat them as ordinary characters, and must not crash on lowercase or on a window with no `C`. + - Pinned by Task 1's self-check cases `anchored("")`, `anchored("nan")` and `near("", "")`. +2. **A cluster whose representative and all members were screened.** The row is dropped, not written with empty lists. OPLM's loader rejects empty member lists. Pinned by Task 3's self-check on `promote()`. +3. **A coarse-level representative that was not screened but is not its own first member.** This means the nesting broke. The build must stop with the IDs, not silently promote. Pinned by Task 3's self-check. +4. **Mutually near CDR3s between `eval_cdrh3` and `test_cdrh3`.** The 85% walk spans both files, so no CDR3 in one file is near a CDR3 in the other. Pinned by the test (Task 4), which checks all pairs across both files. +5. **Proxy rounding.** Coarse folders have few shards, so "1/35 of shards" is not 1/35 of the rows. The proxy must hold `round(rows / 35)` rows exactly, from whole shards plus one sliced shard. Pinned by Task 6's check. + +--- + +### Task 1: The CDR3 matching core (brineylab/datasets) + +**Files:** +- Create: `datasets/oplm-antibody-cdrh3/`, by copying `datasets/oplm-antibody-sequences/` (SKILL.md, `recipe/build`, `recipe/test`, `recipe/params.toml`). +- Modify: `datasets/oplm-antibody-cdrh3/recipe/build`: add the matching helpers below and a `selfcheck()` that runs first on every invocation, before any download. +- Modify: `datasets/oplm-antibody-cdrh3/recipe/params.toml`: the Global Constraints values. + +**Interfaces:** +- **Produces**, in `recipe/build`: + - `levenshtein(a: str, b: str) -> int` + - `near(a: str, b: str) -> bool` + - `anchored(sequence: str) -> list[str]`: distinct candidates, any order. + - `semiglobal_within(query: str, window: str, k: int) -> bool`: Sellers' algorithm. True iff some substring of `window` is within `k` edits of `query`. + - `class PieceIndex`: + - `add(query: str, qid: int) -> None`: splits `query` into `len(query) * 3 // 17 + 1` contiguous pieces of near-equal length and indexes each piece to `qid`; + - `candidates(text: str) -> set[int]`: the qids having any piece as an exact substring of `text`; + - `match(cand: str) -> int | None`: the smallest qid whose query is `near` `cand`, among `candidates(cand)`. + - `selfcheck() -> None` + - `recipe/build --selfcheck` runs `selfcheck()` and exits 0. +- **Pigeonhole guarantee:** if `near(q, c)`, then `c` contains one of `q`'s pieces exactly. `near` allows at most `floor(0.15 * max)` edits, and a candidate within reach is at most 20/17 of `len(q)`, so at most `len(q) * 3 // 17` edits. + +- [ ] **Step 1: Write `selfcheck()` first.** It is assert-based and uses v1's `KNOWN` heavy chain, whose CDR3 is `ARSPYSSSWPGYYYYYYGMDV` (21 residues, so the bound is 4): + ```python + q = "ARSPYSSSWPGYYYYYYGMDV" + assert levenshtein("kitten", "sitting") == 3 + assert near(q, q) and near(q, "ARSPYSSSWPGYYYYYYGMDV"[:-1] + "I") + assert near(q, "GRSPYSASWPGYYAYYYGMDI") # 4 substitutions + assert not near(q, "GRSQYSASWQGYYAYYYGMDI") # 6 substitutions + assert not near("ARDYYYYYGMDV", "ARGGGSDYYYYYGMDVTT") + assert near("", "") and anchored("") == [] and anchored("nan") == [] + assert q in anchored(KNOWN[0]) + w = KNOWN[0][len(KNOWN[0]) // 2:] + assert semiglobal_within(q, w, 5) and not semiglobal_within("WWWWWWWWWW", w, 2) + idx = PieceIndex(); idx.add(q, 7) + assert idx.match(q) == 7 and idx.match("GRSPYSASWPGYYAYYYGMDI") == 7 + assert idx.match("ARDYYYYYGMDV") is None + ``` + Every assertion must hold by the definitions; recheck each edit count by hand before running. Correct a wrong case, never the definition. +- [ ] **Step 2: Run it and see it fail.** + - Run: `podman run --rm -v "$PWD":/w -w /w ghcr.io/brineylab/toolshed/datascience:1.2.0 datasets/oplm-antibody-cdrh3/recipe/build --selfcheck` + - Expected: `NameError` on `levenshtein`. +- [ ] **Step 3: Implement the helpers.** + - `levenshtein`: plain DP. + - `anchored`: `re.finditer("C", window)` start positions × `re.finditer("(?=[WF]G.G)", window)` end positions, keeping lengths 3–40. + - `PieceIndex`: a `dict[str, set[int]]` plus the set of piece lengths, so `candidates` enumerates only substrings of those lengths. +- [ ] **Step 4: Run it again and see it pass.** + - Same command. Expected: exit 0, no output. +- [ ] **Step 5: Commit.** + - `git commit -m "oplm-antibody-cdrh3 starts from oplm-antibody-sequences and matches CDR3s at 85% identity by a pigeonhole index, with a self-check"` + +### Task 2: The heavy CDR3 holdout and the screen + +**Files:** +- Modify: `datasets/oplm-antibody-cdrh3/recipe/build`. + +**Interfaces:** +- **Consumes:** Task 1's helpers, and v1's `stage`, `cdr_mask`, `heldout_sets`, `order`, `connect` (unchanged). +- **Produces:** + - `cdrh3_heldout(work: Path, seed: int, n: int) -> dict[str, list[dict]]`, keyed `eval_cdrh3` and `test_cdrh3`, with rows in v1's held-out schema. + - `.work/heldout.parquet`: one `antiref90` column holding v1's 20,000 IDs plus the new 5,000. + - `screen(work: Path, queries: list[str], cap: float) -> None`: writes `.work/screened.parquet` with columns `antiref99` (string) and `qid` (int32), and prints: + - `screen removed: of (%)` + - `screen cdr3s over 1000: ` + - `screen largest: ` + + If `n / total > cap`, it exits 1 with `screen removed % > %`. + +- [ ] **Step 1: Write `cdrh3_heldout`.** + - Candidates are AntiRef90 representatives not in v1's held-out set, both strata together, walked in `md5(':' || antiref90)` order. + - Number them in pools of 12,000 with the `Anarcii` instance v1 uses. + - Keep a candidate when: + - `chain_type == "H"`; + - `cdr_mask` is not `None`; + - its CDR3 has at least 5 residues; + - `PieceIndex.match(cdr3)` against the CDR3s kept so far is `None`. Add each kept CDR3 to the index. + - The first `n` kept are eval, the next `n` test, with `locus = "IGH"`. + - Print `cdrh3 candidates tried: ` and `cdrh3 rejected near: `. + - If candidates run out, exit 1 as v1 does. +- [ ] **Step 2: Write `screen`.** + - Input: the distinct `antiref99` of manifest rows outside every held-out AntiRef90 cluster, joined to their sequences. Have DuckDB write them as `cpus` parquet parts under `.work/screen_in/`. + - Run a `multiprocessing.Pool(cpus)` over the parts. Each worker: + - builds one `PieceIndex` over the queries; + - for each sequence, takes `anchored(seq)` and matches each candidate; or, with none, runs `semiglobal_within(q, window, len(q) * 3 // 17)` for each `q` in `idx.candidates(window)`; + - writes its `(antiref99, qid)` hits. + - Concatenate the hits into `screened.parquet`, then print the counts and enforce the cap. +- [ ] **Step 3: Wire them into the main flow after v1's held-out stage,** each skipped when its output exists: + - write `eval_cdrh3.parquet` and `test_cdrh3.parquet`; + - extend `heldout.parquet`; + - `screen(work, [cdr3 of every eval_cdrh3 and test_cdrh3 row], params["screen_max"])`. +- [ ] **Step 4: Dry run of stages 1–3 on bcr.** + - Locally set `limit = 3000000`, `heldout = 200`, `cdrh3_heldout = 200`, never committed. + - If v1's `.work/` downloads still exist on bcr, hard-link the two `.gz` files into `/.work/` first; `download()` trusts files of the right size. + - Run: `sbatch --partition=cpu --cpus-per-task=32 --mem=256G --time=03:00:00 --wrap "srun --container-image=ghcr.io#brineylab/toolshed/datascience:1.2.0 --container-mounts=, /datasets/oplm-antibody-cdrh3/recipe/build "`. + - Expected: the self-check passes, then the `cdrh3` and `screen` lines print, with `screen removed` under 5%. The train stage that follows is still v1's until Task 3 replaces it; ignore its output. +- [ ] **Step 5: Measure screen throughput** from the dry run (sequences per second per CPU) and extrapolate to about 316M sequences on 32 CPUs. + - If the extrapolation exceeds 3 hours, move the anchored match into DuckDB: explode pieces and candidates, join on equal strings, then `levenshtein()`. Keep the Python path for unanchored sequences. + - Record the measured rate in the commit message. +- [ ] **Step 6: Commit.** + - `git commit -m "oplm-antibody-cdrh3 holds out 5,000 heavy chains with mutually distinct CDR3s and screens every AntiRef99 train sequence against them"` + +### Task 3: The four train sets + +**Files:** +- Modify: `datasets/oplm-antibody-cdrh3/recipe/build`: replace v1's `train()` and `clusters()`; drop the `antiref99/train` output. + +**Interfaces:** +- **Consumes:** `.work/heldout.parquet` and `.work/screened.parquet` (Task 2). +- **Produces:** + - `clusters(reader: pa.RecordBatchReader, level: str) -> Iterator[pa.RecordBatch]`. Input rows are `(, antiref99, sequence, rep_removed: bool)`, sorted by `order(level), level, (antiref99 = level) DESC, order('antiref99')`. It yields v1's train schema. + - `promote(level_id: str, first_member: str, rep_removed: bool) -> str`: returns the row's `sequence_id`, or exits 1 with ` representative is not among its own AntiRef99 members` when `not rep_removed and first_member != level_id`. + - `train(work: Path, shard_rows: int) -> None`: writes `antiref{90,92,94,96}/train/`, then prints `rows /train: ` and `members /train: mean `. It exits 1 if the four member totals differ. The build checks this as it writes, because recounting about 300M member IDs is too slow for the test. + +- [ ] **Step 1: Extend `selfcheck()`:** + ```python + assert promote("A", "A", False) == "A" + assert promote("A", "B", True) == "B" + try: promote("A", "B", False); raise AssertionError("nesting break passed") + except SystemExit: pass + ``` +- [ ] **Step 2: Run `--selfcheck` and see it fail** (`NameError: promote`). +- [ ] **Step 3: Implement.** + - Kept rows are manifest rows outside `heldout.parquet` and whose `antiref99` is not in `screened.parquet`. + - `rep_removed` is ` IN (SELECT antiref99 FROM screened)`, plus the v1 held-out exclusion. + - A cluster with no kept rows produces nothing, so dropped clusters vanish without special-casing. + - Use one sort and a streaming group per level, as v1's comment requires (an ordered `list()` spilled 90 GB). +- [ ] **Step 4: Run `--selfcheck`;** expect exit 0. Then rerun the Task 2 dry run. + - Expected: four folders, and four `members` lines with the same total. +- [ ] **Step 5: Commit.** + - `git commit -m "oplm-antibody-cdrh3 writes AntiRef90, 92, 94 and 96 clusters of the screened AntiRef99 sequences, promoting a surviving member when the screen took the representative"` + +### Task 4: The test + +**Files:** +- Modify: `datasets/oplm-antibody-cdrh3/recipe/test`. + +**Interfaces:** +- **Consumes:** the output layout in Global Constraints. Its `near` and anchored candidates are reimplemented independently here; it must not import the build. +- **Produces:** exit 0 iff the output is a version, printing one line per failure and `ok` last. + +- [ ] **Step 1: Extend v1's test** with these checks: + - `HELDOUT` is v1's four files plus `eval_cdrh3` and `test_cdrh3`, with `cdrh3_heldout` rows each. + - `TRAIN_COLUMNS` covers the four `antiref{90,92,94,96}` folders, with v1's member checks unchanged. A promoted row's `sequence_id` and `sequence` are its first member's, so `member_ids[0] == sequence_id` still holds. + - v1's four files match the Global Constraints sha256 values. + - Every `*_cdrh3` row is `IGH`, with a CDR3 of at least 5 residues. + - No two CDR3s across `eval_cdrh3` and `test_cdrh3` are near. Check in DuckDB over a table `c(id, cdr3)` of all 5,000: + ```sql + SELECT a.id, b.id FROM c a JOIN c b ON a.id < b.id + AND abs(length(a.cdr3) - length(b.cdr3)) <= greatest(length(a.cdr3), length(b.cdr3)) * 3 // 20 + WHERE levenshtein(a.cdr3, b.cdr3) <= greatest(length(a.cdr3), length(b.cdr3)) * 3 // 20 + ``` + Expect zero rows. Integer `* 3 // 20` is `floor(0.15 * x)`. + - **The screen holds on a sample.** Take the first 1,000 rows of shard 0 of each folder. For each member, take its anchored candidates (regex in the test), cross-join them in DuckDB with the 5,000 held-out CDR3s, and apply the same `levenshtein` predicate. Expect zero rows. Members with no anchored candidate are checked with a semi-global edit distance (a small Sellers function in the test) within `len(q) * 3 // 17`. + - Train row floors: `FLOOR = {level: 0 ...}` until Task 5 sets them. +- [ ] **Step 2: Run the test on the Task 3 dry-run output.** + - Expected: `ok`, except the sha256 check. A `limit` dry run holds different v1 files, so the sha256 check is skipped when `params["limit"]` is non-zero; print `skip sha256 (limit)`. +- [ ] **Step 3: Mutation checks**, each restored afterwards: + - overwrite one `test_cdrh3` row's `sequence` and `cdr_mask_aa` with an `eval_cdrh3` row's; expect the near-pairs check to fail; + - append a row to shard 0 of `antiref92/train` whose member is the `eval_cdrh3` heavy chain itself; expect the screen check to fail. +- [ ] **Step 4: Commit.** + - `git commit -m "oplm-antibody-cdrh3's test checks v1's files by hash, the CDR3 holdout's 85% separation, and the screen on sampled members"` + +### Task 5: Full build on bcr, skill, record, PR + +**Files:** +- Modify: `datasets/oplm-antibody-cdrh3/recipe/test` (`FLOOR`). +- Modify: `datasets/oplm-antibody-cdrh3/SKILL.md`. +- Create: `results/oplm-antibody-cdrh3/1.0.0.txt`. + +- [ ] **Step 1: Run the full build by hand on bcr** with the committed `params.toml`, using Task 2's sbatch line with `--time=12:00:00`. + - Expected: `ok` from the test, except the floors, which are still 0. + - Record per-stage wall time, `sacct -j -o MaxRSS,Elapsed`, `.work/` disk and output size. + - If the screen's tripwire fires, stop and bring the `screen` lines to the human. Do not raise `screen_max`. +- [ ] **Step 2: Set `FLOOR`** to `int(0.9 * rows)` per folder from the printed `rows` lines, and rerun the test. Expected: `ok`. +- [ ] **Step 3: Write `SKILL.md`** with v1's sections: + - **Frontmatter:** as in Global Constraints. `description` says the dataset is a heavy-chain CDR3-identity holdout at 85% plus four AntiRef cluster levels with AntiRef99 members, for OPLM's antibody anneal. + - **Files:** the six held-out files and four folders. State that a row's `sequence` is AntiRef's representative unless the screen removed it. + - **Reading it:** always `sample_members: true`. + - **Recipe:** the spec's §5, with the parameters. + - **The test.** + - **Compute:** Step 1's measurements. + - **Gotchas:** + - the D3 exception: four levels in one dataset, by the PI's decision of 2026-11-12; + - light chains are held out by AntiRef90 cluster only; + - the screen is CDR3-only, so V-gene and framework similarity remain; + - no screen against uniref70 or deepclust70; + - the screen over-removes slightly by design. +- [ ] **Step 4: Check.** + - Run: `python3 -I scripts/check.py datasets/oplm-antibody-cdrh3`. Expected: exit 0. +- [ ] **Step 5: Push, record and commit (ask the human first).** + - In `datascience:1.2.0` on bcr, push: `python3 -I scripts/push oplm-antibody-cdrh3 `. + - Write the record: `scripts/record oplm-antibody-cdrh3 0 "sbatch --partition=cpu --cpus-per-task=32 --mem=256G --time=12:00:00" bcr.scripps.edu < `. + - `git commit -m "oplm-antibody-cdrh3 1.0.0's record: its candidate passed on bcr"` +- [ ] **Step 6: Open the PR (ask the human first).** + - The branch carries the design commits, this plan and the dataset. The PR body links briney/oplm#78. + - A reviewer other than the builder approves; then `gh pr merge --merge`, pull `main`, and `scripts/promote oplm-antibody-cdrh3 1.0.0`. + - Expected: tag `v1.0.0` on `brineylab/oplm-antibody-cdrh3`. + +### Task 6: Proxy data and the standalone baselines (CoreWeave) + +**Files** (cluster, not git; back up before editing): +- Create: `~/projects/oplm/mup_redo/ab_proxy.py` +- Modify: `~/projects/oplm/mup_redo/ab_cdr3_eval/cdr3match.py`: take the shard folder and the output prefix as arguments. + +**Interfaces:** +- **Produces:** + - `python ab_proxy.py `: + - symlinks whole shards in order; + - writes the remainder as one sliced shard, so the destination holds exactly `round(rows / divisor)` rows; + - prints `rows `, then `residues_per_row `. That's the mean over rows of the mean member length, the same quantity as `overfit.sh`'s `115.6`, which v1's shard 0 gave as 115.62. + - Proxies live at `/mnt/data/briney/oplm-data/proxy-1of35/cdrh3/antiref{90,92,94,96}/train`. + +- [ ] **Step 1: Download the dataset.** + - Run: `snapshot_download("brineylab/oplm-antibody-cdrh3", repo_type="dataset", revision="v1.0.0")` with `~/venvs/oplm-a140e98/bin/python`, into `/mnt/data/briney/oplm-data/hf/oplm-antibody-cdrh3`. + - Check the file count and sizes against the Hub. +- [ ] **Step 2: Write `ab_proxy.py`** with an assert-based check at the bottom (`if __name__ == "__main__" and sys.argv[1] == "--selfcheck"`): build a 2.5-shard toy folder of 10-row shards in a temp dir, proxy it with divisor 2, and assert 12 rows (`round(25 / 2)`) with every `sequence_id` drawn from the first rows in order. + - Run `--selfcheck`; expect `ok`. Then build the four proxies and keep the printed `rows` and `residues_per_row`. +- [ ] **Step 3: Score the baselines on `eval_cdrh3` standalone.** + - Use `ab_cdr3_eval/run_eval_split.py` with `SPLITS=eval_cdrh3` and the data dir set to the new dataset. + - Checkpoints: `decay140k-u25/output/checkpoint-175000` (protein-only control) and `decay140k-u25-ab1.5/output/checkpoint-175000` (old v1 train set). + - Run it as a 1-GPU sbatch modelled on `ab_cdr3_seen_unseen.sbatch`. `admin_override` is allowed for this small job under the standing rule. + - Expected: JSON lines with `loss_heavy`, `accuracy_heavy`. + +### Task 7: The sweep (CoreWeave) + +**Files:** +- Modify: `~/projects/oplm/mup_redo/overfit.sh` (back up as `overfit.sh.bak-pre-cdrh3`). + +**Interfaces:** +- **Consumes:** Task 6's proxy folders and their `residues_per_row`. +- **Produces:** `AB=1.5 AB_SET=antiref90 bash overfit.sh decay 140000 25` creates run `overfit-decay140k-u25-ab1.5-ar90` (`-ar92`, `-ar94`, `-ar96` likewise). Without `AB_SET`, behaviour is byte-for-byte as today. + +- [ ] **Step 1: Edit `overfit.sh`:** + - `ABH3=/mnt/data/briney/oplm-data/hf/oplm-antibody-cdrh3`. + - In `decay`, when `AB_SET` is set: + - skip the v1 symlinking; + - add `-ar${AB_SET#antiref}` to the name; + - pass the proxy path and its `residues_per_row` to `write_config`. + - In `write_config`, the antibody source key stays `antiref96`, so W&B keys match the earlier runs: + - its `path` is the proxy folder; + - its row weight uses the passed `residues_per_row` in place of `115.6`; + - when the extra argument is set, add eval `antibody_cdrh3` with `path: $ABH3/eval_cdrh3.parquet`, `type: antibody_cdr3`, `every: {steps: 1000}`, `metrics: [loss_heavy, accuracy_heavy, perplexity_heavy]`. +- [ ] **Step 2: Dry-run all four** with `DRY=1`. + - Diff each `config.yaml` against `decay140k-u25-ab1.5/config.yaml`. Expected differences, and only these: + - the antiref96 `path`; + - the three row fractions, which change slightly with `residues_per_row`; + - the new eval entry; + - names. +- [ ] **Step 3: Ask the PI to approve `admin_override`** for four 4-node, about 10-hour runs. Then submit the four and move each to `admin_override` with `sudo -n scontrol update JobId= QOS=admin_override`. +- [ ] **Step 4: Check the start of each run.** + - Check the first eval at step 141k in W&B: `eval/antibody_cdrh3/loss_heavy` must exist. + - Grep each log for `Traceback`. Expected: none. + +### Task 8: Analysis and report (briney/oplm#78) + +- [ ] **Step 1: Pull the step-175k results** for the four runs from W&B: every `eval/*` key and P@L. Put them beside the control and old ab1.5 numbers from Task 6. +- [ ] **Step 2: Seen/unseen split on v1's eval files**, per run. + - Run `cdr3match.py ` for each of the four proxies. + - Score each run's checkpoint on its own seen and unseen files with `run_eval_split.py`, as one 1-GPU job. +- [ ] **Step 3: Apply the spec's decision rule (§8):** + - prefer the lowest `eval_cdrh3` heavy loss, provided v1-file unseen loss is no worse; + - flag any protein CE more than 0.002 off the control rather than choosing it. +- [ ] **Step 4: Post the tables and the recommendation to #78** (`board comment 78`), naming the W&B run IDs, the dataset tag and the jobs. Leave the production choice to the PI. +- [ ] **Step 5: Update memory `antibody-cdr3-eval.md`** with the runs, the result, and the `AB_SET` usage. diff --git a/docs/superpowers/specs/2026-11-12-oplm-antibody-cdrh3-design.md b/docs/superpowers/specs/2026-11-12-oplm-antibody-cdrh3-design.md new file mode 100644 index 0000000..972d117 --- /dev/null +++ b/docs/superpowers/specs/2026-11-12-oplm-antibody-cdrh3-design.md @@ -0,0 +1,175 @@ +# `oplm-antibody-cdrh3`: Design + +Status: draft 2026-11-12, from a design conversation with the lab's PI. Tracked on briney/oplm#78, the anneal-mix experiment. It builds on `oplm-antibody-sequences` v1.0.0 (design `2026-11-09-oplm-antibody-sequences-design.md` on that dataset's branch; skill `datasets/oplm-antibody-sequences/SKILL.md`). Calls made in writing that nobody discussed are marked **(call)**. + +## 1. Why + +oplm#78 tested a decay (anneal) mix with antibodies on OPLM's 1/35 overfit proxy (170M model, 140k→175k from the stable checkpoint). It used AntiRef96 clusters with AntiRef99 members at 1.5%, 3% and 6% of tokens. Results: briney/oplm#78 issuecomment-6812930334. The PI chose **1.5% for production**. Two findings set up this follow-up. + +- **The current eval sets reward recall.** v1 holds out whole AntiRef90 clusters, nothing more. Of eval_gte2's CDR3s, an exact copy is in the training shards for 27% of heavy chains and about 85% of light chains. +- **Extra antibody data buys memorization faster than generalization.** Going from 1.5% to 6%: + - loss on unseen heavy CDR3s fell only 1.50 → 1.41; + - loss on seen ones fell 1.31 → 0.79; + - the seen-vs-unseen gap widened 0.19 → 0.62. + + The proxy repeats its 6M antibody chains 3–12 times, and production would too. + +So the question now is which antibody **train set** generalizes best at the fixed 1.5% share. Coarser clusters (AntiRef90, 92, 94 representatives) with member sampling from AntiRef99 spread draws across more distinct families and repeat exact sequences less. Answering this needs an eval where "unseen" means something: a **heavy-chain CDR3-identity holdout**. + +## 2. Decisions (the PI's) + +**Revised 2026-11-12, after measuring the screen.** Scored against about 7,900 real heavy CDR3s, a 24,000-sequence train sample showed the screen at 80% would remove about 10% of train at 10,000 held out, against a 3% tripwire. Removal scales mostly with the number held out, not with CDR3 length. The PI chose **85% CDR3 identity at any length** (same-length-only matching saved little once at 85%) and **2,500 + 2,500** held-out heavy chains, with the tripwire at **5%** (about 4.3% expected). Conserved Cys104 and Trp118 were already outside the compared CDR3 (IMGT 105–117). Below, every threshold, size and bound reflects this revision. + +**Revised 2026-11-13, after the first full build.** The screen removed 5.134% of AntiRef99 train sequences (16,222,603 of 315,971,320), over the 5% tripwire. Removal is concentrated in common CDR3s: 1,472 held-out CDR3s each remove more than 1,000 sequences, the largest 358,021. The PI kept the held-out sets as drawn and raised the tripwire to **6%**. + +- **Holdout: CDR3 identity, from AntiRef, first.** A donor holdout needs an OAS rebuild with donor annotations, and waits. +- **Heavy chains only, at 85% CDR3 identity.** No heavy-chain CDR3 left in training is 85% or more identical to a held-out heavy chain's CDR3. Light chains stay on v1's AntiRef90-cluster holdout. Their CDR3s are low-diversity: holding them out by CDR3 identity would gut light-chain training or leave almost nothing unseen. +- **Train sets to compare, all at 1.5% of decay tokens on the 1/35 proxy:** + - AntiRef90 representatives with AntiRef99 members; + - AntiRef92 → 99; + - AntiRef94 → 99; + - AntiRef96 → 99, rebuilt under the new holdout as the baseline. +- **How a winner is chosen:** + - lower CDR3 loss on unseen CDR3s; + - lower heavy CDR3 loss on the new CDR3-identity holdout. +- **The paired-chain axis of #78 is dropped.** +- **One dataset, four clustering levels: v1's exception to `docs/DESIGN.md` D3, extended (PI, 2026-11-12).** All four train sets have to share one holdout and one screen, and one recipe guarantees that. The skill states the exception, as v1's does. + +## 3. Goals, non-goals, anti-goals + +**Goals.** +- A new held-out heavy-chain set whose CDR3s have no 85%-identical CDR3 anywhere in training. OPLM's existing `antibody_cdr3` eval reads it unchanged. +- Four train sets sharing one holdout, so the four decays differ only in the train set. +- v1's eval and test files carry over byte-identical, so earlier #78 numbers stay comparable. + +**Non-goals.** +- A donor holdout, newer OAS data, paired chains, or an SHM mask: these need the OAS rebuild. +- A CDR3 holdout for light chains. +- A leakage screen against uniref70 or deepclust70, as in v1. +- An AntiRef99-representatives train set: the sweep doesn't use one **(call)**. + +**Anti-goals.** +- A screen that is only as good as a regex. A held-out CDR3's near-duplicate must not slip into train because a CDR3 boundary was misplaced (§5.3). +- Changing OPLM's code for this. The CDR3 eval, the loader and `sample_members` already handle everything here. + +## 4. Source and identity + +Same source as v1: AntiRef v2022.12.14, the Zenodo manifest and FASTA, with the same MD5s and the same nested linclust levels (100, 99, 98, 96, 94, 92, 90). + +~~This is a **new dataset**, not a v1 version. Name: `oplm-antibody-cdrh3`.~~ **Revised 2026-11-13 (PI):** this is **`oplm-antibody-sequences` 2.0.0**, a new version of the same dataset: the same data with a different eval and test holdout. Against D3 by the PI's decision; the skill says so. 2.0.0 drops 1.0.0's `antiref99/train` (PI). Train results are not comparable across the two versions. + +## 5. Steps + +### 5.1 v1's holdout, unchanged + +Run v1's hold-out step with v1's parameters (`seed = 1`, `heldout = 5000`). The 20,000 held-out AntiRef90 clusters and their four files come out identical to v1.0.0. The test checks that by hash against v1.0.0's files. + +### 5.2 The heavy CDR3 holdout + +- **Candidates.** The representatives of the remaining AntiRef90 clusters, walked in a seeded order (`md5(':')`, as v1) with a new seed, `cdrh3_seed = 2` **(call)**. Number each with ANARCII (IMGT), as v1 does. +- **Keep a candidate when:** + - its chain is H; + - every residue is numbered; + - its CDR3 (IMGT 105–117) has at least 5 residues **(call)**; + - its CDR3 is less than 85% identical to every CDR3 already kept. Without this, one family could dominate the eval **(call)**. +- **Split.** The first 2,500 kept are `eval_cdrh3.parquet`, the next 2,500 `test_cdrh3.parquet`. Columns are v1's: `sequence_id`, `sequence`, `locus` (all `IGH`), `cdr_mask_aa`. +- **Identity.** Two CDR3s `a` and `b` are 85% identical when `1 − levenshtein(a, b) / max(len(a), len(b)) ≥ 0.85`, that is, when the edit distance is at most `floor(0.15 × max length)`. Length differences count against identity **(call)**. + +### 5.3 The train screen + +Remove from train: +- every sequence in a held-out AntiRef90 cluster (5.1 and 5.2), as v1 does; +- every sequence whose CDR3 region holds a near match to a held-out heavy CDR3. + +**What "near match" means here.** The match is CDR3 against CDR3: the PI's 85% is CDR3 identity, not containment. The guarantee must still not depend on locating each train sequence's CDR3 exactly. ANARCII on about 300M sequences is too slow for CPU, and a regex (`C(...)[WF]G.G`) agreed with ANARCII's CDR3 on only 93% of eval chains. So the screen tries every plausible boundary pair. + +- **Window.** For each train sequence, the window is its C-terminal half, `sequence[len(sequence) // 2:]`. CDR3 starts near residue 105 of about 120, so the half always holds it, and the screen never needs to locate Cys104. Planning chose this over a window that starts 10 residues before Cys104: it is a superset of that window, so it can only remove more **(call)**. +- **Anchored candidates (call).** A candidate CDR3 is any substring of the window that starts right after a `C` and ends right before a `[WF]G.G` motif, 3 to 40 residues long. + - Each candidate is compared end to end with each held-out CDR3 by §5.2's identity. + - A train sequence goes when any of its candidates is 85% or more identical to any held-out CDR3. + - This tries every boundary pair, so a misplaced boundary can't hide a match. It also can't match a held-out CDR3 against a fragment of a longer train CDR3, which keeps shared germline D or J stretches (`…YYYYYGMDV`) from removing unrelated antibodies. +- **Unanchored sequences.** A sequence with no anchored candidate (about 6% of v1 train sequences by the regex) gets the conservative fallback. It goes when any window substring meets the identity bound. The check runs as a semi-global edit distance of at most `floor(3 × len(query) / 17)`. That is a superset of the 85% bound for any substring length, since the bound allows at most `floor(0.15 × max length)` edits and a substring more than 20/17 of the query's length is out of reach **(call)**. +- **Light chains.** The screen covers every train sequence, light chains included. A light chain rarely matches a heavy CDR3, and skipping chain typing keeps the screen simple **(call)**. +- **Speed.** How to run 5,000 queries against about 300M sequences fast is the builder's call. For example, by pigeonhole: a match within edit distance `k` shares an exact `⌊L/(k+1)⌋`-mer with the query, so an exact k-mer index narrows the candidates before alignment. + +**The build prints:** +- how many sequences the screen removed; +- how many held-out CDR3s removed more than 1,000 sequences, and the largest removal. + +A public heavy CDR3 drawn into the holdout takes its whole family out of train. That is the point, and the count makes it visible. + +- **No per-CDR3 cap.** A cap would bias the eval away from public CDR3s **(call)**. +- **Tripwire.** If the screen removes more than 6% of AntiRef99 train sequences, the build stops so a person can look before anything publishes **(call)**. + +### 5.4 Train sets + +From what remains, write four folders in v1's antiref96 format: +- `antiref90/train/`, `antiref92/train/`, `antiref94/train/`, `antiref96/train/`; +- each row is a cluster at that level: the representative's `sequence_id` and `sequence`, then `member_ids` and `member_sequences` listing the distinct AntiRef99 representatives inside it, the representative first; +- the same seeded shuffle, 1,000,000-row shards, row groups of 10,000, and zstd as v1. + +Every folder holds the same set of AntiRef99 sequences, grouped differently. So the four train sets differ only in how training draws from them: a uniform cluster, then a uniform member. + +The nested clustering makes each representative its own AntiRef99 member, as v1 checks for AntiRef96. If linclust's nesting breaks that at a coarser level, the build stops. + +**The screen will remove some representatives while members of their clusters survive.** CDR3 identity doesn't follow linclust's full-length clusters. Equal member sets across the four folders matter more than AntiRef's choice of representative, so: +- the surviving member first in the seeded order (`md5(':')`) becomes the row's representative: its `sequence_id`, its `sequence`, and the first entry of the member lists; +- a cluster with no surviving member is dropped; +- the skill says that in this dataset, a row's `sequence` is AntiRef's representative unless the screen removed it; +- OPLM always reads these folders with `sample_members: true` **(call)**. + +### 5.5 Print + +v1's `rows` and `locus` lines, plus: +- the screen's counts from 5.3; +- each train folder's rows and mean members per row. + +## 6. The test + +v1's checks, plus: + +- v1's four eval and test files are byte-identical (sha256) to v1.0.0's; +- `eval_cdrh3` and `test_cdrh3` each have 2,500 rows, all `IGH`, and every CDR3 at least 5 residues; +- no `sequence_id` appears in two held-out files, or in a held-out file and a sampled train row's members; +- **the screen holds.** For a sample of train rows (several thousand members per folder), no window is within the 85% bound of any held-out heavy CDR3. The test reruns the semi-global check on that sample; the build's own index is not trusted; +- all four folders' member sets have the same total count; +- each sampled row's first member is itself; +- the screen removed no more than 6% of AntiRef99 train sequences. + +## 7. Compute + +Build on bcr, as v1 (Zenodo is slow to CoreWeave). v1's download in `.work/` can be reused if it's still there. + +My guesses, to be measured: +- v1's 37 minutes; +- the screen, a few CPU-hours with a k-mer index; +- four train folders, about 25 GB each, about 100 GB of output. + +## 8. How OPLM uses it (briney/oplm#78) + +This is in scope. The plan builds the dataset first (a PR here, cooked on bcr), then runs the sweep and reports on #78 **(call)**. + +**The runs.** +- Four decay branches with `overfit.sh` (`AB=1.5`), one per train folder, from the 140k stable checkpoint to 175k, 4 nodes each. +- **The 1/35 subsample** for each folder is the run of whole shards, in shard order, whose total member count comes closest to 1/35 of that folder's members. Repetition depends on distinct AntiRef99 sequences, not cluster rows, so this keeps the four runs' pass counts about equal; the realized fraction is reported **(call)**. +- Same recipe as `decay140k-u25-ab1.5`: uniref70 at 25% of tokens, antibody tokens taken from deepclust70. +- Row fractions come from each folder's mean member length, as `overfit.sh` computes now. +- `admin_override` for these runs is to be confirmed with the PI at launch. + +**The evals**, every 1,000 steps: +- `antibody_cdr3` on v1's `eval_gte2` and `eval_singletons`; +- `antibody_cdr3` on `eval_cdrh3`, with metrics limited to `{loss,accuracy,perplexity}_heavy`. Its light metrics would be empty. + +**Scored standalone on `eval_cdrh3`:** +- the protein-only control, `decay140k-u25` at 175k; +- the old `decay140k-u25-ab1.5`, whose training never excluded these CDR3s' neighbours. Comparing it with the rebuilt AntiRef96 run shows how much the old numbers were inflated by leakage **(call)**. + +**Report on #78**, at 175k, for the four runs and the control: +- heavy CDR3 loss on `eval_cdrh3`, the primary metric; +- v1-file CDR3 loss, split into seen and unseen CDR3s with `cdr3match.py`. Run it per train folder, since each holds a different 1/35 subsample; +- protein CE on deepclust70 and uniref70, eval and train probes; +- P@L on cameo, casp, pdb40 and pdb40_2048. + +**The decision rule:** +- Prefer the train set with the lowest `eval_cdrh3` heavy loss, provided its unseen-CDR3 loss on v1's files is no worse. +- Protein CE must stay within the ±0.002 spread seen between protein-only schedule variants. If it doesn't, flag it rather than pick that set. diff --git a/results/oplm-antibody-sequences/2.0.0.txt b/results/oplm-antibody-sequences/2.0.0.txt new file mode 100644 index 0000000..c262901 --- /dev/null +++ b/results/oplm-antibody-sequences/2.0.0.txt @@ -0,0 +1,754 @@ +--- +dataset: oplm-antibody-sequences +version: 2.0.0 +hub: brineylab/oplm-antibody-sequences +branch: candidate-716e0c9cff7d07e266f94824249eaf7500078164 +commit: 08de241378cb9ce06370926d52cf811a78dbdc6c +tree: 716e0c9cff7d07e266f94824249eaf7500078164 +tools: datascience@1.2.0 +runner: flyloft job submit --cluster bcr --cpus 32 --mem 256G --time 12:00:00 --tool datascience@1.2.0 +host: bcr.scripps.edu +started: 2026-11-12T18:50:09Z +exit: 0 +--- +screen removed: 16222603 of 315971320 (5.134%) +screen cdr3s over 1000: 1472 +screen largest: 358021 +rows antiref90/train: 152351629 +members antiref90/train: 299748717 mean 1.97 +rows antiref92/train: 161836404 +members antiref92/train: 299748717 mean 1.85 +rows antiref94/train: 175083828 +members antiref94/train: 299748717 mean 1.71 +rows antiref96/train: 199443463 +members antiref96/train: 299748717 mean 1.50 +rows eval_singletons.parquet: 5000 +locus eval_singletons.parquet IGH: 2216 +locus eval_singletons.parquet IGK: 1897 +locus eval_singletons.parquet IGL: 887 +rows eval_gte2.parquet: 5000 +locus eval_gte2.parquet IGH: 2875 +locus eval_gte2.parquet IGK: 1305 +locus eval_gte2.parquet IGL: 820 +rows test_singletons.parquet: 5000 +locus test_singletons.parquet IGH: 2228 +locus test_singletons.parquet IGK: 1911 +locus test_singletons.parquet IGL: 861 +rows test_gte2.parquet: 5000 +locus test_gte2.parquet IGH: 2801 +locus test_gte2.parquet IGK: 1377 +locus test_gte2.parquet IGL: 822 +rows eval_cdrh3.parquet: 2500 +locus eval_cdrh3.parquet IGH: 2500 +rows test_cdrh3.parquet: 2500 +locus test_cdrh3.parquet IGH: 2500 +eval_singletons.parquet 5000 rows +eval_gte2.parquet 5000 rows +test_singletons.parquet 5000 rows +test_gte2.parquet 5000 rows +eval_cdrh3.parquet 2500 rows +test_cdrh3.parquet 2500 rows +antiref90/train 153 shards 152351629 rows +antiref92/train 162 shards 161836404 rows +antiref94/train 176 shards 175083828 rows +antiref96/train 200 shards 199443463 rows +screen sample: 6734 members, 7281 anchored candidates, 200 unanchored checked +ok +file antiref90/train/shard-00000.parquet 129402674 5c99169220da53c1faced042b2cedca0e691b491928c1783f57b32760da5f542 +file antiref90/train/shard-00001.parquet 129879632 7ede76ec87d88149760ed20780340cba3ca8e5908bff1a77ff256c95df92ea7d +file antiref90/train/shard-00002.parquet 129848224 5bdce7e12b04ca0c7ee4e64ed1a186bf6f88dc2892548b11cd7b13992a082e51 +file antiref90/train/shard-00003.parquet 130249098 dceb77b6b3c78d65ad655ce3924dddf24123f422d5ed55f0269b823775dfa9d9 +file antiref90/train/shard-00004.parquet 130287928 15973ecf9df094af97e24b290e8f52526791eca858e9783dced633e9ceee93b8 +file antiref90/train/shard-00005.parquet 129763452 6492af9bfe37cb323fc0f43f05216963db32ac62bd625c231799a749a72309e7 +file antiref90/train/shard-00006.parquet 128933521 37897998f0871dbbe45a9c3fe7ec12b75e9b702672515110e16b8174187a6152 +file antiref90/train/shard-00007.parquet 129850719 b1fd04992c6d25b1b73883ad34e868e0e5ffb6b357d049f7328a66ac57c49bfc +file antiref90/train/shard-00008.parquet 129890288 d9af44bf14a69f8815f9f5963f6d95034f3fd0b342627cb0d1d20265b8586f25 +file antiref90/train/shard-00009.parquet 130213609 5ef64cde7c3fc2fe4a420f9e4becc36ebca9aa8db3b499340e6c9647df1a3776 +file antiref90/train/shard-00010.parquet 128927825 6b678488dad7a9593e7af8a5e904ff7911e6d5e50d307817b4ef9e6848029769 +file antiref90/train/shard-00011.parquet 129105247 655187a824e80cb0aa8f36fa79343a294ca1dcd821282b00f31116d1cdf7d83b +file antiref90/train/shard-00012.parquet 129805072 dc21afa3d039c01d1b9536d260b44c3a02de6afafc936f2dc8e10783ad5d3d62 +file antiref90/train/shard-00013.parquet 129933305 7885b536068e7566bf88a843c9b23cb0a409a852e0789049140b13c13ddd1dd9 +file antiref90/train/shard-00014.parquet 130732002 e2bd4061a0dbd7b20b8e433b312409c83f3811cccbc40bb07a84e6c044af961a +file antiref90/train/shard-00015.parquet 129324454 6f981e6cf57572882b67ac2f5d90ff8b988c0649ca56a03ca472b5bf0264cad6 +file antiref90/train/shard-00016.parquet 129711388 666705bf4f7fcb7196746822622667bd41f514bda37ec305272042eb078cc91e +file antiref90/train/shard-00017.parquet 130581496 d3effc1c31fdf2544293b71b89916d7dcce17e744235dbe6e64264e609ce2853 +file antiref90/train/shard-00018.parquet 129638071 1abdcfd16fcd0d7ffaade725a7ae3ab64f4a688b256a3f68aa4c95db5603c399 +file antiref90/train/shard-00019.parquet 129266003 4f199bf9ec93f84137a35dc77b7c2fec402e8efa8d0f82969fd269a66c06740b +file antiref90/train/shard-00020.parquet 129557298 ac8f8acc8c479111bc08ead1334752ac372a20571998cf5c2b27ab1db15f5a6e +file antiref90/train/shard-00021.parquet 129611118 58cc199210a0a85be872b8d0285fbfdd37c49314e9ca5dc593a0ca31a562de3d +file antiref90/train/shard-00022.parquet 130660263 165efa8524613acc79ab9b0448b94fcebfddb1ee4e8b2ec03dbf5bbcf6dbc79a +file antiref90/train/shard-00023.parquet 128813007 77f37eb82b59a4ea2176c84edfa813fc9eaec2e15e45a956b4011b25fe43a0ee +file antiref90/train/shard-00024.parquet 129942682 bf7cdfa3ee01aaeb43df34a72ade91193798745b61139cba162390460ae8b7ce +file antiref90/train/shard-00025.parquet 129966904 357cbb77bbb9ae5b5482559a78966a800a9565a17a23a9f3500860baa3e965de +file antiref90/train/shard-00026.parquet 129977564 e22df95e57bef289c4047b8b44dc5a88d718755b97250cb0df41804b7cddcfd5 +file antiref90/train/shard-00027.parquet 130932325 942203bd3e93145fc8c5a7f1ad9a8ca2d19eeb9c5a9affc3aaf58f37e27fab54 +file antiref90/train/shard-00028.parquet 129973445 13bc2627ebc19ee7e65d086b335d2ee2ecae7b7346214fc52dc3a3eee6577236 +file antiref90/train/shard-00029.parquet 130080105 b91f965c0ca99dcd061c54b96ccc1c5e24069be7953d3bf03258a8a3c1c72e89 +file antiref90/train/shard-00030.parquet 129280992 21c34b5b23d4ecb70f2a83d2f13d59c2f2fc1e61cb571578e5c6e457a294bf01 +file antiref90/train/shard-00031.parquet 129440449 938f402736bf4b710159c314eb193178530dde1274e382f3a5984e7cfc6c3ed7 +file antiref90/train/shard-00032.parquet 129470674 d3af53e02338e407223afae70c39c34302da4a5c12997157d595849a4e19de47 +file antiref90/train/shard-00033.parquet 130431332 44c179f40ee768d1b2e86eba4af6da6b657fe5ca0c18e88de55df3fe587e5eda +file antiref90/train/shard-00034.parquet 130287476 613d4dcb49fc029a0a7e083b508f68d41b972cb71e2923401e1432ae737288b2 +file antiref90/train/shard-00035.parquet 128413840 39addf7bc9f87def6c22d46ddd14188e0102875562b8ed63d651e25f0226cd9a +file antiref90/train/shard-00036.parquet 129069976 23a38e939eee0411f83ffb08943878c753eed497739e526320a77558fb16ba01 +file antiref90/train/shard-00037.parquet 130270287 523b602b21b1e5a655c50869929933429f094741dc516aee94cfaf487895b4e6 +file antiref90/train/shard-00038.parquet 131762892 d952d4596387e986e946b6e8e6ae90e5eda215ecf78de42731fed19dbed524e9 +file antiref90/train/shard-00039.parquet 130298718 ec8e2511f68e13a699a6acd6d454475556897cb09f20a657cde8a2d9e31c4515 +file antiref90/train/shard-00040.parquet 129697271 355a95dd2d2c924e4d1cbb4e4e6042b6d9693954ae39a7f30a836d8740e0a5e2 +file antiref90/train/shard-00041.parquet 129797976 cc5620bf8269daa568621948136355379353b79f6b202c9eebec74dc9fcaa47a +file antiref90/train/shard-00042.parquet 130464769 faa52f5c81964c5c0db74fbfc0cdc5172e831076e55cc48494e0645d3c6ea04b +file antiref90/train/shard-00043.parquet 129680110 e97252dfcd5daa066f73536bbb0ffdbfad3ed5a75ff31342e073fe7c5ac1e313 +file antiref90/train/shard-00044.parquet 129171048 b498a7effa9e2f46ea01e7aa1624d8ac332bb2885217c550a5bf7b1e4da408c9 +file antiref90/train/shard-00045.parquet 130418335 77cfabfd521906331999dfa817a0085b488a29abd49c3e3b4e425e56cc264551 +file antiref90/train/shard-00046.parquet 130478927 00a39b91b6efa64a7923df41bad74ff4c352699c45e12255e207c8d5db19fb53 +file antiref90/train/shard-00047.parquet 130365954 8e48ede15db0b5f23e50717b7f4ae329b94dc7631acfbcbdcaa766f9f087182a +file antiref90/train/shard-00048.parquet 130655861 a597e6cd6ec98a8c7d30ee4df4d963f92449756500ee90f620b25bffaabeb9b7 +file antiref90/train/shard-00049.parquet 129049789 def064f38fac70e16f83c86c72970e2784c6647321ed13118731224d4ad7c7ff +file antiref90/train/shard-00050.parquet 129794492 4199675e3cd6d6d298022a6bf9f2e6dec00f4cbe251be56fb3101b524dd63fc7 +file antiref90/train/shard-00051.parquet 129445655 386c437343f5faa29128565d7480f195614e77dd7c59c9acd46adfa00255cfbc +file antiref90/train/shard-00052.parquet 129804114 74d50f6de1df36af5c5b431d26b614789fa2169ef5579f7fc2b2afc069a34a9f +file antiref90/train/shard-00053.parquet 129390869 6ff5a239616364531de3494fb8b0bde6cb4eefc3291e91c7daf55c77e7000f05 +file antiref90/train/shard-00054.parquet 130989021 f27e93a34941cbf6314f79067575c7b8c798f1289f20d70425361a84443dccf3 +file antiref90/train/shard-00055.parquet 129915857 efa91bade01d5b6d1590b9db47634ab8d8c07008c4188911e085f61a3ced932c +file antiref90/train/shard-00056.parquet 129109970 475db123301512fa200b0471d8af7f537863f623467abb5b9430297488aaf589 +file antiref90/train/shard-00057.parquet 130308612 7700c1a12430504a387ca52908e4b355292e23fe8f0e382b8914782e793220b4 +file antiref90/train/shard-00058.parquet 129072062 9283524ffc5b1d6517a11de0563abb9bbe12f83eebe812f9c3b263e8979c97c8 +file antiref90/train/shard-00059.parquet 130129783 0547c37247f3128f9b085681d27a1b29555dc1b1b2f51f419a8c9830e39f1117 +file antiref90/train/shard-00060.parquet 129850707 4022dbe33ae52b6991c365f05e9f62704133df0019d25ec817988a46c55d1357 +file antiref90/train/shard-00061.parquet 128547569 0485f5d674d249d71cc786df5a307bb8f9ddee2fad083ed8a3f7cf566ed51a71 +file antiref90/train/shard-00062.parquet 131442704 6e61fa63e2baed41310e0770f7397a38e25bf038c5d5305ee458f95de585a822 +file antiref90/train/shard-00063.parquet 129092049 5ef1ca751a97e8247b691601bab2a2393449f885d9a4f0a7d7ea5d75782a071f +file antiref90/train/shard-00064.parquet 130482957 9461fbe59c950a9828ec328a83b600cf84a45ac4fb17af53bfd5b8e2710e69ec +file antiref90/train/shard-00065.parquet 129802425 6acee67adfbb5cc7496eba333de2a02939f86bf65b56897d7ac06b30561e616a +file antiref90/train/shard-00066.parquet 129045277 516efecfd64bd6b27d262b727559364cf2a58601fc6aa4a1fd2e46040b42c5e7 +file antiref90/train/shard-00067.parquet 130742018 d4f6b76187d6fb74b7eb5b86f896d940de6dd8d7649363552da82ab656b70066 +file antiref90/train/shard-00068.parquet 129631265 400d35f219fc8b3f7626cd922883efb7c4890ccf8e00cb11d95c2398096a7179 +file antiref90/train/shard-00069.parquet 128639356 30214155aa50dd92bffe7410ad6900091c50cad52a8e4277d184c09d334887f4 +file antiref90/train/shard-00070.parquet 131175520 81c0a77724d47498269d8b72bab9c4c038bc345034b60f3c5357e7418e1a56c6 +file antiref90/train/shard-00071.parquet 129255588 4d984e051cf1cbff976a5017c0888b5d26c24b7a2450ed54ee26526b8d9cb1fc +file antiref90/train/shard-00072.parquet 130647510 e780be5804a1ef56b51c1a98f5ab905894291a7d20df1846c9ffb2b4971f9aaf +file antiref90/train/shard-00073.parquet 129821163 da5308c8c37bc083645a78c189ec11e4eb5a2ce340367634166ad0824d9b3ff6 +file antiref90/train/shard-00074.parquet 131174701 4af73aa1751f58f5126517d72d14038f01f93f04e26b33f8a8c0b48e5ea0486d +file antiref90/train/shard-00075.parquet 129958917 23bab27fa9a87126cb92c2f78d49bcd09748aa6487e22fd673e250ad8c0ac0a6 +file antiref90/train/shard-00076.parquet 130726892 47b9f12285cdadf1eb8533efb9f557faca2ae5d3e315261ec7ca941064e4026c +file antiref90/train/shard-00077.parquet 130816557 32a4a521cc09833d8caeea1092cd0afbe6e71344bdfa441a570108f14b037445 +file antiref90/train/shard-00078.parquet 128964369 35aa526ea5171d56a9c5d09c806fa8f3b92090f14f1c66d67cc38da94ba7e364 +file antiref90/train/shard-00079.parquet 131282402 680d3c0855e2668d7971ec7552e01be7ad19d7971f7f559adb58a517e8db8c84 +file antiref90/train/shard-00080.parquet 130011855 dcf528711e37fb48ae2a60f647eff94a812742e0dd86040c5153c8c8b2f0a79e +file antiref90/train/shard-00081.parquet 130458492 17970363eefec16f1ad38635bb46ee003114cd0c69f30a4cecba21ed5e2ac835 +file antiref90/train/shard-00082.parquet 130144434 41bb7c14de9020236a2a6dd4f2b1a71f462965c9e7433e2f10d36de4c5e4d81e +file antiref90/train/shard-00083.parquet 129767310 b936e614f3354e6684d25c8a17691402bff2066bbfade609cec6e28afcb7fa77 +file antiref90/train/shard-00084.parquet 129313662 9b8874556f26f7677cfb04f547bcc8e1b7cb86c50b7787f834366bcf3fde25b2 +file antiref90/train/shard-00085.parquet 130285611 adfab4200642ea1ba7dfdd2611bc19df640e2224d33e9f37aa641834dadff453 +file antiref90/train/shard-00086.parquet 129308287 75efc29d660d3320450ace5377496d743a70c2d84c4abc9db604cc26dc34ad45 +file antiref90/train/shard-00087.parquet 129199043 c4860b51d96a8bc77f1be96e246e02970660de2d5412905affb2ef5a94313aa2 +file antiref90/train/shard-00088.parquet 131111548 1b60c60ccb47c98d870829e4eb20dcb938e1b65a7223c863cc37c7e1033fe2c7 +file antiref90/train/shard-00089.parquet 130171953 f5dd00c9ae61392aa75096eb23c217e0c8d9a896544a53f23c74f22661462b56 +file antiref90/train/shard-00090.parquet 129414018 7539a577d399a28e1bf813ec2ab8903e1f88f36309c1a9bcde6b8d1bf9097767 +file antiref90/train/shard-00091.parquet 130687095 8d3909d0b5511a502585f6e70840de93709d7f48ca2a6d0071ef01ed18b60c8b +file antiref90/train/shard-00092.parquet 130341247 a6fa66250f2c00fb5cddb50301e0456fe617b86323360de320c852c8a3e1b77c +file antiref90/train/shard-00093.parquet 130362078 c5698ae1eb1f171f3895533811c7e3584cda19231b30849d2cd1ffb088ab062b +file antiref90/train/shard-00094.parquet 128906958 9f4acaed7aaad4cf130acd03bfd121c099ca3c28e81460ef9da8f30ef0a30eb7 +file antiref90/train/shard-00095.parquet 129734146 668cbdac905e790e751c811905b752df291c8252db63ddcb6f846083dd1b3491 +file antiref90/train/shard-00096.parquet 130048339 d2ef938290567f269a2f0530538b3b3d7d1cb974a34193331c3925126eb776bc +file antiref90/train/shard-00097.parquet 129554296 4d8a917e4dfc286462c8026c040c14f53d515796b9bfb65bf575d42dcd82890d +file antiref90/train/shard-00098.parquet 129178988 a74d7da03aa14aff9ad782bf3c01742c6a852d929b6bf17581a84f1874e3558c +file antiref90/train/shard-00099.parquet 129676571 86d848df2dd314552c7bd732fa28599e5cfafdb2901d18143ca0c78b8d726a07 +file antiref90/train/shard-00100.parquet 128809708 92c53b827b4f97c32fb63666fe8e406b82058a59c8092b8070c5679448705419 +file antiref90/train/shard-00101.parquet 131200616 8bb6e22cecb5d5ee3c2cdcccca3b1a937db8c244520a9f68abf198367f82a085 +file antiref90/train/shard-00102.parquet 130421829 d5f64583cb6641919e829e2fc53876e6fa6e8db918b8854f8c8013b1af413687 +file antiref90/train/shard-00103.parquet 130708520 203158a7c54ea099b057c569acc0f60276f975efa37116d3732cf2627c43c3dc +file antiref90/train/shard-00104.parquet 130485484 88814756addd60a70fc41841cac50ba0e6712ebab0d851157eb1186e38acf187 +file antiref90/train/shard-00105.parquet 129113188 1d16e6591127ab4ecb2a268c1d3364cfccc719cf9287026a9489462fec0987df +file antiref90/train/shard-00106.parquet 130547755 f061572222011b369fe380409e0994fddb630b113ce5ca23eb388ccb99518b6d +file antiref90/train/shard-00107.parquet 130529781 f3c4ce73090287beaa7440a7bc7fbb5c9458471657cd5802367dc2d9d43565e8 +file antiref90/train/shard-00108.parquet 130679744 8dce3804e212a371ba54b7b0dda1f231674f0f1c3d66e43283b25694ce99bf76 +file antiref90/train/shard-00109.parquet 129965482 6a5c06695b986c2bc27f4e615295951007ed65f13634c3cea5a92bdf04b33c00 +file antiref90/train/shard-00110.parquet 130023163 3ce5576bcbb244befd0ac47a708ea8a4db53bb2798eff534c46dd7a55400d9b7 +file antiref90/train/shard-00111.parquet 129451621 8e7b27d479245334ce344f610e9c5b5c37a6e5f785d43dc80aa93eeea5d74151 +file antiref90/train/shard-00112.parquet 130890368 4642b1714ed07b77231757240538090875e2fa18173ee5ddd8fbf72dca19a36d +file antiref90/train/shard-00113.parquet 129333179 d3df9d8ca13a0550b85ffae5733e9a7d0cef9cd3a2943657291d98928412ee4f +file antiref90/train/shard-00114.parquet 131202310 9187ad721159783ad4606451a0b8706af4331e99af2a82ec35ca0310064191f4 +file antiref90/train/shard-00115.parquet 128961325 db9e506196d916f3b6373ef6366a7469b526e437cbf546fdc46b86bad8a64927 +file antiref90/train/shard-00116.parquet 128799833 703c1fa93ce43b53fa4ea8850a060e10bcdad57f594ea396204aa73109baeaf8 +file antiref90/train/shard-00117.parquet 130185738 e93d20b85428268755857810282a77dcf3f141fa3a3e972fa7d9c3539b0d726a +file antiref90/train/shard-00118.parquet 129847652 b4b63ac022585b305092f8a6428e237c796b7b3e1b2060f057bfb90306b3abca +file antiref90/train/shard-00119.parquet 128849055 30c41e6a2a7d8ddd4571939bb980f0637bbd5321c84fd025b24986d0b354cb2a +file antiref90/train/shard-00120.parquet 130297155 6773834bea2b6edbe9b2ad91db0167b3ff992e4bd6ffc5b4f557635452da6800 +file antiref90/train/shard-00121.parquet 130210855 e183187203c403a647ffefd5a034a1987243be4b19d8586dc7da6b90882027dc +file antiref90/train/shard-00122.parquet 129944162 cb87119f5dc85c768537f602fc47051957b8d9425c0f6a433210ff101605db11 +file antiref90/train/shard-00123.parquet 130895175 57d88a6149f1f0eee4f2b6596480b7a69c9adfe2cbd9781b665b83e7af9b1bd8 +file antiref90/train/shard-00124.parquet 129413359 12f797631fd6e9bf0e07d655b56126d249255768b18589de6ae7e2ceb6a7df48 +file antiref90/train/shard-00125.parquet 130146229 a2a8153004e1abab6bad18d47e31a2ec4fb576cd08ad0b01b6cff2ec8cab17e1 +file antiref90/train/shard-00126.parquet 129866954 e77a2796b8b6d49bcd154c89acade5f014ebc03d67b09ddbeb77f9b778912de2 +file antiref90/train/shard-00127.parquet 129809005 459851d9eb614cb7baaa23b1f82ca5973327478db9e350235f3c5bcef7801467 +file antiref90/train/shard-00128.parquet 128266828 5307b819f8af1bb422a86adbd3e0d622be1408899f689b80207496ff8355f63f +file antiref90/train/shard-00129.parquet 130352671 172866b82dd9f3b4e9141b40ebd736ae20e9212f0912d4365296583d498f83c1 +file antiref90/train/shard-00130.parquet 131435142 563dfd8485a92d7f6e619eece9926d3f780fe2d455e58fd5c7e211b62b655257 +file antiref90/train/shard-00131.parquet 131197059 9081321a42e527e323458b70f48ff63822c9beaa5a19b30a5df593e8052e9dfe +file antiref90/train/shard-00132.parquet 130085289 9765ec4d6b4c807adbb8acdaef19abdb5399300010f29a82ecbd708a4b357e3f +file antiref90/train/shard-00133.parquet 130170121 1d0ff5a7aea56d85b6a13a653df52804bfa7fd0809945eeed3fe1b2019be5949 +file antiref90/train/shard-00134.parquet 129019874 87179a41e9aa76f161acb4620257d7f9a1387c734328deb1ff0ff63b871545da +file antiref90/train/shard-00135.parquet 130602942 b0a8738616f013bc41c07b08de90100967a53d190579b811d91d837fec7c575f +file antiref90/train/shard-00136.parquet 130010790 bbbb0a9f96cb055e1e7e16d1ce373dc0d143f68de6b7df76dbe3f3de3b2d5721 +file antiref90/train/shard-00137.parquet 131358016 b23ea643c37b34aa7792b49e1742eabd5531e6095608c38d498fa9b9f14c64f6 +file antiref90/train/shard-00138.parquet 129885316 868c5f50dc3630a40349e2f5f93127346f9cbc4de6a17ec17bec16da59a27f94 +file antiref90/train/shard-00139.parquet 130178316 ffeeb845d31c674816b379b0f4539e1db37d7dc476464f4b00ed1c16808bb922 +file antiref90/train/shard-00140.parquet 128783169 bb1c6f67810fd4eb4d2c3dbd034b77c30ef0454e20b9cb18d9fb0227166e04fe +file antiref90/train/shard-00141.parquet 130258940 c08a9df41426f0751725abbdad3a76962650124f5160e6fcbc842a83eaf088e4 +file antiref90/train/shard-00142.parquet 129627015 9d91da4ebb04c73285f126436b52b8835bd8adbb113d496860d0a894bbecc06e +file antiref90/train/shard-00143.parquet 129974412 b2a664a5b23cf2f18d0215b56ac998b6236cde42bed7499b6e6a0b02f68b75d5 +file antiref90/train/shard-00144.parquet 129515842 70e2289becb83489e4ba00409b85e95aa002add471f4e10cef9e842691098686 +file antiref90/train/shard-00145.parquet 129803042 c194a3075d57c15f6bd2d6cbcda350bd6fa66ad732a21a2941c1b44a21578cbf +file antiref90/train/shard-00146.parquet 129549094 d83d25f6978717dddc707355265c2ec298142e60a0c02e8a0e98103bf4ef504e +file antiref90/train/shard-00147.parquet 128868059 fd33c69aab0450bb9f80a0b4434a68f461f78225c21992639982bdead3f7469e +file antiref90/train/shard-00148.parquet 129945277 ebe4e9f8c6cfedeb5ca61f611069a0197cb9fd2f1708c34c3afcb2ade24bc4c6 +file antiref90/train/shard-00149.parquet 130714512 ef3073babd0537e54436eb985fe1bce9c740e05854f63a5abd6627f8caa4d182 +file antiref90/train/shard-00150.parquet 130388014 62253e87fe628d9c8c71b1a777893df738a767c88fecfc06390431ff96a493d5 +file antiref90/train/shard-00151.parquet 128782996 8ea5d4b81fc41076d51a77187b7e575ae38cd58843948997d5923d53917964e3 +file antiref90/train/shard-00152.parquet 45628599 587a82d1d8897910add26ed2ace02468a20abac46b6e231766c2812dce456126 +file antiref92/train/shard-00000.parquet 126886925 6bf89c9e0eaed10b4859b577164a75acd89885370118ddec49c225a10b5e2921 +file antiref92/train/shard-00001.parquet 127043201 fcd8205fbe6f43bd594b33e012808b6c225ab87ff779bd43dbf3042d01d1c41c +file antiref92/train/shard-00002.parquet 126308376 405649622492e0d031ccf36b36239169e6d55d88ed48c5632c88209d14afa406 +file antiref92/train/shard-00003.parquet 127086740 e0c46bcf13d8c541b0fb355fd70ccca6173284f4ab88d3fbf3c7088578cbb7bf +file antiref92/train/shard-00004.parquet 127157825 a3fb99067729b2de24c0a617ae8dd16d33ad23c34c64ffd87ff9f22fec404778 +file antiref92/train/shard-00005.parquet 126816563 233419dc649876c1e1773e073c7aebec4bb5979e33b495c6b18a4151c318fffc +file antiref92/train/shard-00006.parquet 126069858 8d72a8eb41c22c88ff5c8b6eab988766e55c8f4bd265b2ae316fa09018affddc +file antiref92/train/shard-00007.parquet 126902641 ada0ca8bc185eb187eb2763ed9406a97a827bdc1cd91e3cf8de26d235b4e14ea +file antiref92/train/shard-00008.parquet 125920213 dafb15e153b61261a1af0c1d7a8e98140c9da2a55784c2031f6e60ed11714f2b +file antiref92/train/shard-00009.parquet 127401102 722d685266772b5c13d98b95be2cbc0e100a858a8b68c8634eb52dfdaafba3a8 +file antiref92/train/shard-00010.parquet 126392044 e6b38e2463191c98b5e0e288aae43e293ce8a042ef61f945cfaa251933d7bb04 +file antiref92/train/shard-00011.parquet 125835244 d7ad544489b891c81a9d793e784412e6b3cb996ff361c7c40bdc57730f1645e0 +file antiref92/train/shard-00012.parquet 126650375 62efc7eb9b1e8e00438e28ce3c34b42f669230435a4fdc24ce30e8e95c4e82d8 +file antiref92/train/shard-00013.parquet 126977362 dda45b59bffe762e99ef8e8677ae02dbb4ac64fe8312e2856652eedd4a44cc61 +file antiref92/train/shard-00014.parquet 126423764 dad58f38fd656d8ccce54ec0f4fbd54892069f18cfdfaeca3462208382df4a8d +file antiref92/train/shard-00015.parquet 127734091 970ba2fe2532b58fb08c0cb691dd071f8f3965cab0a6615db74af52583400be6 +file antiref92/train/shard-00016.parquet 126340701 1a85a543a4e08b07f16e6a80b6a0e03fd1809ba0bce89faa9293eeb062e83633 +file antiref92/train/shard-00017.parquet 126603467 5e7b091c84f2311725b1edc3d87befdf16ee2352312f0776feb3ce7d4a3f5a73 +file antiref92/train/shard-00018.parquet 127215982 2133b0fac9e5db41b283d77d06d83b445a29443a2dfd55f1de52f3a9479ccd9c +file antiref92/train/shard-00019.parquet 127152178 b3a7aca20798041b2fa5b0f5b5d1a4837af651eab0e2c7686187f671259a7a94 +file antiref92/train/shard-00020.parquet 125742697 d4ec0e4d48e5fdd210eb5ec576e278e4c88add5eae4b7244bb52262c2203b172 +file antiref92/train/shard-00021.parquet 126768933 5a1a03f8e1004ce549bb0327a221d17f8ed8297722cfb0afe9dc417609209170 +file antiref92/train/shard-00022.parquet 126914261 5d7695143534dee40c4cc011b1f7a5f8e5c4506b24b791df4f5e4fc551796378 +file antiref92/train/shard-00023.parquet 127610156 7b133aa224ca82092c4a1c43d0d39d7d95a64d2126b098e3acab7e0167f1b747 +file antiref92/train/shard-00024.parquet 126335792 5b1c695959aaa1652f3efbf06ec0fc52d1719ed1e90e714b95ffc18d7060bfc3 +file antiref92/train/shard-00025.parquet 127255273 aa8b4738fe50b46223199fd90bb0cbb07f9fd5fd9802e1a474fd0b9dfe3af9ba +file antiref92/train/shard-00026.parquet 126851533 04f4cb1a9f6e4eed7a218dfe8a5b4b4f6610efb1ff13b9fd607213e3fe1ce90b +file antiref92/train/shard-00027.parquet 126513402 36e29f6bdeb5c066afb454eabecf4fa182971e38b1e0de5b955c6b9994616971 +file antiref92/train/shard-00028.parquet 126976590 6d56e51851ea0c97b865513e702864414ce6a4cf3d2f808a1a8f83eb19a0de28 +file antiref92/train/shard-00029.parquet 127840960 ad71d8e809b480a7a7b6678ca80042f8f584fa05b92bbf9b7ed965d435ca2c45 +file antiref92/train/shard-00030.parquet 126950390 5c999d937e777d34843f22ca85792e1233952ee0fbb533d9b4b44a2dd7c1e27d +file antiref92/train/shard-00031.parquet 127335308 09dc12c9b791bf218a0daa89401137e9538a220064ad9936a187f1231dbc8c78 +file antiref92/train/shard-00032.parquet 126766669 6158b03ae1fb7745cf7523b29d21ab85c210cf56f0fc02a5bb648845bbca1fe4 +file antiref92/train/shard-00033.parquet 126774020 c0a7fa039991468481a8942fb77af6a66f89b5d670bfe201dddab97d8fd985e7 +file antiref92/train/shard-00034.parquet 126666168 4902780935979532784ae414b671c82f1e5af8acfa9d1332d0f0b3da117777b0 +file antiref92/train/shard-00035.parquet 127927160 430776729e1235fde1a87d9d1a746ad314338a4a96ded78444b0495c836ef766 +file antiref92/train/shard-00036.parquet 127549331 066287fc2cb5aabe31042d1be1d00fffbfe8269a91dec415eaa95a7146bc05d6 +file antiref92/train/shard-00037.parquet 125355226 438311ff20c19be875e4ff622c2f1083c1e2115f24d51b72e459dca9d828fb5e +file antiref92/train/shard-00038.parquet 126338549 6a8dd5efeab50e3997ce08b415d9096c80eaec69a0f21e7834b916f71cea2375 +file antiref92/train/shard-00039.parquet 126813899 a3c10693826b30ff652e71e6a90ced30754381e8658001f3a32950bbcfb4f1e4 +file antiref92/train/shard-00040.parquet 126808949 af2dad4d2954daf1d56217fc7b00242e47a3d7ce0a58738f17f21d1892b42296 +file antiref92/train/shard-00041.parquet 128906373 e5952f9184787ec45b58c57a5bd81023a39a350df967d322e12831561867ca91 +file antiref92/train/shard-00042.parquet 127137670 f391612c86a1da9d3897d4abf8b98457fec85a5e386f6434d31bf738e10e1388 +file antiref92/train/shard-00043.parquet 127355394 60cb820980506fa43373b14c640c017520bc5c086b788460a5889665cc5e35a0 +file antiref92/train/shard-00044.parquet 127555475 89aa0539e4f841b73057953d80b72caaf665ffefad493b395c4987796a7cc373 +file antiref92/train/shard-00045.parquet 127146386 b07cd94542d648b2b75a02b1472be7ab12ee89ef85fbe2fab084e281e6f19ac1 +file antiref92/train/shard-00046.parquet 126107823 f09ae95c5e9551ff3e96990e5ac628cbf5c8df7abdbf2d3d1b824a1cc31caa95 +file antiref92/train/shard-00047.parquet 126425849 f2511f89a508513dc609a5058ef64f892eb12f50cfd3b7a07074a67e104ef448 +file antiref92/train/shard-00048.parquet 127229048 2cd5791e445d55c1ff0339c3eb3dd9078258dad20e4376f8f7bd85265928f1ce +file antiref92/train/shard-00049.parquet 127518225 17260b920aacc97eb67265772f263f1b2927dd14ced25603938ae7b91d896954 +file antiref92/train/shard-00050.parquet 127032387 0d5fb8722c27cbab687950de64e8890a4ddf59baf2a101374d2cb92579e092ff +file antiref92/train/shard-00051.parquet 127496676 59cdcd98e50cb66f5d3dc4209847635b560d6c7006cd898ef6b1118d57960c38 +file antiref92/train/shard-00052.parquet 126855092 e466371ce93f1d189b8a700942750551357b67d49ffec13364f62c46aa188a0e +file antiref92/train/shard-00053.parquet 127375253 3fe26d94d5f6efe9b046a5b08a44d072cc2acfbf9113cc164e6b10275ef24b67 +file antiref92/train/shard-00054.parquet 127067848 43ef00779d98b33cba6946d9b47542e1434a3b8be6e0e26c2191b5f445961d2c +file antiref92/train/shard-00055.parquet 126734389 7eb3959a518b28153da7c2c7d8222a38751ed0a4f7afe2218b8d127fea971080 +file antiref92/train/shard-00056.parquet 125752305 6d2d9b2b330407f968da2655756ed8014bf6fe82d321d9cc77a47c9b4426bd4c +file antiref92/train/shard-00057.parquet 127734382 86d9b4c0f4c0ab82befefd15e6657b990c86f9220996f444fb5214ea8525006f +file antiref92/train/shard-00058.parquet 127267235 ec5ae85ba5d96bee160f6dd0fb3f058b8991a6177003ecfc76a33b9ff9551f7c +file antiref92/train/shard-00059.parquet 127182400 83c2c633a1c05955f1d9cab0a6837a304f6d996dace0d1f7eb347db99b13475c +file antiref92/train/shard-00060.parquet 127407697 0160e9e2b0bf4dcc8c81234d9c22a5ee928544214a482896bd6416850e56e8b9 +file antiref92/train/shard-00061.parquet 126701762 a37a1899e04bf3b232c8c8b038fc1906ba99fd0cfd4375934292e4663f6be10a +file antiref92/train/shard-00062.parquet 126795906 8190f2bd8feeb44609ba2253ae98ce8af3dc7c6fe9bfd6e25ee1d25f5352038c +file antiref92/train/shard-00063.parquet 126379597 8aa9117734843de789dd0360171cd47683bc3eba606c29cd9a0a6dd2cb0a9137 +file antiref92/train/shard-00064.parquet 126711603 7ef4c069e9dacd5c3ac0d8e7165eb00db6eee09c8426fa3a4d4e02afbc1c5fa2 +file antiref92/train/shard-00065.parquet 126416673 e8244467aae9793d3e7bcb50c58b13cda71e2081bf68df9edb314b0a253c4514 +file antiref92/train/shard-00066.parquet 128438454 7771cd8435a66e9873e35017c4cbdcf8c9e16aaa023532e102111d830f5adb04 +file antiref92/train/shard-00067.parquet 126081860 d26054c98672d383679895a56fe16e2c798e9e4a8b6e50d6d8514f5219763142 +file antiref92/train/shard-00068.parquet 127091349 9748538bdf2d19c93d8dc2b2a598c873e8111bfda222d3aa251151c8300cfd34 +file antiref92/train/shard-00069.parquet 126976748 ced1c8e5f37fa448040602c2d9b4e1caf859ef5791c26eb265913567260e58a9 +file antiref92/train/shard-00070.parquet 126281024 460f2213849c9dde2a5302aac5aff80aff6bb704bbcaa9ab63ffc162d2b0e405 +file antiref92/train/shard-00071.parquet 127026675 b60e8af00a25762b1f73014f28bb943c16f1c1b9fcfc7d8d6f86b81e3fafead1 +file antiref92/train/shard-00072.parquet 128136872 a0473f08440b0ab29f6bbe17aff367645e9d97d21bba1fe54b8e7f5f60cb0644 +file antiref92/train/shard-00073.parquet 125985461 ce7f17e716623d23be0c6ceaf28e34e0ff2d4c5ecf0de59fa1aff357ee1d1d3e +file antiref92/train/shard-00074.parquet 127676798 9aa5a4fc06123c8867bf7533c6120e4e817d865402ace15f330cb38760a5d21f +file antiref92/train/shard-00075.parquet 126456217 2346178d83e4039022ad35919e69cd3b258c4853ee5b798839aaf0d35324203b +file antiref92/train/shard-00076.parquet 126537996 b01ddfc10b99ecd03d307386aab6e36732acf3b46e0148e4ebf954390024e5f1 +file antiref92/train/shard-00077.parquet 127733281 0cf3fc7c4add5b6b7afff091d76d39a9aa1e26d81513aa2138151f45562d5823 +file antiref92/train/shard-00078.parquet 128006981 9197b1758d03a50e2ee37874a8b3bb4f0327814bdae7335eb5f5715b257cf8ff +file antiref92/train/shard-00079.parquet 127445189 e6f4a0aab3e9de19c33867974f24d9a2231a69e66acf13a0bbe93e86de4dd451 +file antiref92/train/shard-00080.parquet 125991763 86d212047c5d6aa43c512531fb55a06eede357b855fb206cc8b51f8503b62e91 +file antiref92/train/shard-00081.parquet 127928343 2dc3345cfa13023bbe64ae28cc147838091f07092e2c13210805ceb504244625 +file antiref92/train/shard-00082.parquet 127085518 7c67f9d64ad310ce4005294057243ea2eaa6dee8c2cda7c7169d15661cc0f2c9 +file antiref92/train/shard-00083.parquet 126072361 2d8ee9fcaac5aaec5834e03b53a71cea8dfa5c0618ee2df4a0d2b41f0c045935 +file antiref92/train/shard-00084.parquet 128002579 f69931992283e9cd04982d79c1f322f5b10132bc54c67b95391785fe4d9dcea0 +file antiref92/train/shard-00085.parquet 127051488 80119f5fbd4024b541fa313f86cebf12ce2df0c3ea4904cfb98425f7f7322791 +file antiref92/train/shard-00086.parquet 127034230 bb0e8c818c033c3100f77862ae9340e9fcca1097ea9c5acb561da4c3274cc180 +file antiref92/train/shard-00087.parquet 127165082 374cb0387e72621a29f548dc68df56e59f36661f8c69238a848ce92e22a6c63c +file antiref92/train/shard-00088.parquet 127302027 0290ed806eff8f0db41a163aef1bec12682dd1764c82b1e31dcbe2cb87d7d829 +file antiref92/train/shard-00089.parquet 126615235 a67904c6b76de50a041de027668121b5c55847e3ceef8e16bea905e1e981e197 +file antiref92/train/shard-00090.parquet 126689345 cf81b375f7f9e5cadf397ef2ebf660d5091b0653d81d6febf368f3c6829f2d2d +file antiref92/train/shard-00091.parquet 126342686 0752214518e6d425ec7dae5eda094b55a1e05bbec1c1488266082f9109976077 +file antiref92/train/shard-00092.parquet 126054266 ee2d3e831a685225f5e9aff6d9b337952ef0d525b083c4703c1ecdb7badb7bf1 +file antiref92/train/shard-00093.parquet 126920061 4622ce5f7647a811071a9370323af15b004ea771adde2e610f23596c38a4a417 +file antiref92/train/shard-00094.parquet 127899329 a2ea052dfec2e473e28637ce75cd1a34aed2ea16f7f2133e43e963799c00cadb +file antiref92/train/shard-00095.parquet 126957628 7dc9d4b94860060664f29300359bae0624a0edb1e324921d4bf66e2731432945 +file antiref92/train/shard-00096.parquet 126637488 e0275122ced374fdac42478e50328abb91430acc77ab28a4414937f7f749855e +file antiref92/train/shard-00097.parquet 127328374 a77cbaef1b4101024a2c5dcaeded5d8e32650f8ceb03745bc8e400e3b420bda6 +file antiref92/train/shard-00098.parquet 127250668 635d9397036c0839ef482e271a04a95d2bf6131987337c0a460edb6432357b30 +file antiref92/train/shard-00099.parquet 127223485 cc3239c8391c9d397e99cd8c38217d50b01a6e2b720fc81687d3873a69efd780 +file antiref92/train/shard-00100.parquet 126113728 c5faed16bb67ca619b6b2b8377d6b5faf00185804f6d6824fdde905affe9c447 +file antiref92/train/shard-00101.parquet 126964250 8df08a8c3418f6211a7a8e4db20e26d678b86134b915246829ee7053a98275a7 +file antiref92/train/shard-00102.parquet 127149801 d867e2d1fb0357db22f410984f787612f635405c3bca9a3378225d666443e60b +file antiref92/train/shard-00103.parquet 126320462 ef828d517009571fd911b76a23640a054f8831b450b1fad433691d72512eec09 +file antiref92/train/shard-00104.parquet 126301570 520dac3b4f2fa41c0b81ff5659160d573f4d60feace2e47bc3d314d5b5b4f90c +file antiref92/train/shard-00105.parquet 126893810 49f5bfb1423104a221308fcef17ec3240f26463c2eb8bd73320e19bb5161627e +file antiref92/train/shard-00106.parquet 126766362 dd9a4cc00f77ce95c701f190a6e075c152b275f7b7a0325def6ed9682d74d358 +file antiref92/train/shard-00107.parquet 126734944 7b3354530e0277a6f693165e3dd84d02358e9f2683d0ead22bd640bfbaebbfe3 +file antiref92/train/shard-00108.parquet 128195199 8018fff6cdabcffd4ce71b3557c7c035380e92aff8ff32bda94181da369973a7 +file antiref92/train/shard-00109.parquet 126990300 e85870ebb47681d23816295cbcdb5487af91fd34a8992b6964a88b1eeef0ea8c +file antiref92/train/shard-00110.parquet 126773297 45702bf9010dac3b72053414a7c08c46fa37e8047146d3dc25e6f2dc2d971837 +file antiref92/train/shard-00111.parquet 127289414 65f888db1836c4a6e8f8762886a2177b388209d57a04f612a2b7999a2c7dfcfe +file antiref92/train/shard-00112.parquet 126500893 258780cbbd9450bce80e96071d65943ed58366e938e1a18a31309d82e8371ffa +file antiref92/train/shard-00113.parquet 126513082 315b2f1c085a12f8e843cf23435b2b54cf38362d2fcc829d271334fbecd2d078 +file antiref92/train/shard-00114.parquet 128236902 38d952e0da6b42f4749d3270170bc416654c6089c7ffd045d8715d2e5616b6b5 +file antiref92/train/shard-00115.parquet 127602760 b21c92dd220a085b1f7924e51ca337e4aeb99bae7847bce752a5a5b8da11dea6 +file antiref92/train/shard-00116.parquet 126455775 c62c730823caac2737c32b26c09601a1342684cc42c8a30dc38e830b627e550e +file antiref92/train/shard-00117.parquet 127152560 0463ffb7d1dc8ff2ea9df0bc609e3c2333f2a98852f23ad0fae8fe6a448f0409 +file antiref92/train/shard-00118.parquet 126587479 e9f75ce7ade4f874caef06dd2901a3223b674698cd6e8eadf201af001f88e359 +file antiref92/train/shard-00119.parquet 127676916 52682359f4dcdaf9f3b35d0d520fa4d7125067c9f973591b5c2693d8c27e15ff +file antiref92/train/shard-00120.parquet 126506284 25e08bad6b6c002ecb6059a9093adbd06a702253f064f26c00abd7b2ff4cf033 +file antiref92/train/shard-00121.parquet 127833859 b707f4a38195e490838c46bf033fe7dd4136fec7aaa7e3f39e2775465073131f +file antiref92/train/shard-00122.parquet 126009761 837896ea73e13060070cb850cea40af19cfbf50027e43f03a864dac7ab672b05 +file antiref92/train/shard-00123.parquet 126275251 107fded23a81f6399745c5919092df7f49820a306f2e611c5aa8339c01cbafb0 +file antiref92/train/shard-00124.parquet 126631109 ee43408715d7f2c3fc40ec6b7599d880a7c036735bec434a3366fa83feaab43a +file antiref92/train/shard-00125.parquet 127232726 3d99d4f540487a9d98edc6b3d45e838498d356ddfb1047a6d230d11c6161e933 +file antiref92/train/shard-00126.parquet 126011779 aac2fedde15d74f836f0301b6c74c71db93d110799cebf656d0e308a0e0bdc36 +file antiref92/train/shard-00127.parquet 126805952 3758cda8a1261d0d7e805ad4a69912ef04fc8cdc6847f559b7a5092c4e32a9c7 +file antiref92/train/shard-00128.parquet 127088624 37f8a63fd723d3f93cc1e1005d92ba975269c350465a02b86f5389fd1876f2ba +file antiref92/train/shard-00129.parquet 127407137 708338d3362dcec3fc6270397aef6f69cb0e9bbec79558ef0e175ecf1cc63c4b +file antiref92/train/shard-00130.parquet 127092167 104c61d43d840ce89cd56ac416f65eafe8376544599bf4377046189383665cef +file antiref92/train/shard-00131.parquet 127241233 9f9480aa4e8d6f17e6f1c6b3570867494f852f891a23079f071ac0cbcf564ece +file antiref92/train/shard-00132.parquet 126292118 5e0e0b7e9104f47fb53be69f1c9a94f5e2b17476442650199868053c56817461 +file antiref92/train/shard-00133.parquet 127550685 692977c9e70795b2ce1a8dfb5726c3e281f29f78ce33eaac143080dc2a767592 +file antiref92/train/shard-00134.parquet 126829563 5ce373b466f1bf0398fd8fe2ba511de3220eb05e1b0be42bcc68cb0a310c4619 +file antiref92/train/shard-00135.parquet 126584257 4283694544ec76626c8b197cb4d9e255be779997bc9fbe17f08e95eb6c4aed15 +file antiref92/train/shard-00136.parquet 125778619 0752a886e4b0d7a517145d91d403d33d30b33d9be26a1f5b15a62573fe366620 +file antiref92/train/shard-00137.parquet 126919321 3da87396a89a7cd305176eeecdca5538e778f5dd8a40eaf227c772299774f9ed +file antiref92/train/shard-00138.parquet 128222321 d4c6f6810c2e25be9e4ee8b0d99cba46c5d2c2ee195ecbc8dc5738fdf179ac06 +file antiref92/train/shard-00139.parquet 128456801 b3034f313d13629b9b21f0e37a12696c9d953e46c91624d5b4c1c1dc0eab6b16 +file antiref92/train/shard-00140.parquet 126969620 0004e88c92b121188cf1fed9f12e082cc34f14aeff1692621aeea3877fb19c1d +file antiref92/train/shard-00141.parquet 127550616 7ebdae6db06725a20ba3cad4d74770beec472f43105efd5c09e08b1eefd8a5c3 +file antiref92/train/shard-00142.parquet 126530812 5a55cddceb7e2bf0384ffdb77918568cc5a3923ed2599b4c5b9d81817c9c7e31 +file antiref92/train/shard-00143.parquet 126366746 2635cf1e399d5acf41ab4ca15e77e9b0d0977dced7499fd8a18b669bbb334700 +file antiref92/train/shard-00144.parquet 128319444 9107c6d9562453ff30a29b9c73f4738cff3e1e5f98f4560bc1c6ae3ae09cf5c6 +file antiref92/train/shard-00145.parquet 127057424 a70fedc4e194b756d32de51f48f40341a8d51b337af151f3ac6604dd1e2d7892 +file antiref92/train/shard-00146.parquet 127291186 bcef12519c380bbb2935eb2466cdbeff96cc4445de623aaf9510d5de674c7767 +file antiref92/train/shard-00147.parquet 126250884 0b0736786b5383d8c4345900e251d046aa6fa2e87da0f445b75d2fbd1802640a +file antiref92/train/shard-00148.parquet 127179936 a97d42962c35b2cde25550db169810122044aaa168d7c8a653befdd526164392 +file antiref92/train/shard-00149.parquet 125893892 e4dc8c5a188671cf5cd563fbe332770d7f5098e603fb3f8e4ed1e5a522810264 +file antiref92/train/shard-00150.parquet 127877767 a0e59501d016a11244a02d8a8bc27f67a19517bd5402b03e2a33d530aa85e8da +file antiref92/train/shard-00151.parquet 127362330 35d1bc1ce926e31b70d16f9161f68bb5dc633cad7e733b6d828bc2e469a45b61 +file antiref92/train/shard-00152.parquet 127412302 6c6c596a32e654bcc3cf9f81fc1f7c242c1a133cd62a12c85fc94ab5e5528713 +file antiref92/train/shard-00153.parquet 126374146 3030c34548777ab09e9f2d72a0192258d2952a3b54c0f3f14d1ba8278592ac90 +file antiref92/train/shard-00154.parquet 126648765 7cf6f597f059ba029b732f1d1cf983764e97c4e3c0f6466255d933dee8f71be4 +file antiref92/train/shard-00155.parquet 127027816 610298aca7b31a7e8c73c440bc0dee60a68a3797bd8190bbbfe6f411750197d3 +file antiref92/train/shard-00156.parquet 126822680 94a2a2f85d3c581c54b11411c5cd8239ff68aea13f86e90a569eb96e6ae83b31 +file antiref92/train/shard-00157.parquet 127207529 fec4780e2288ae913b72faee8613393ab1dd59e9a098ed41c99b54db686e91b0 +file antiref92/train/shard-00158.parquet 127546749 b7050f6158852f81b5e9135cfabc1852e0a4180bccff4bff3c247a8b9a2b9acd +file antiref92/train/shard-00159.parquet 126879108 4815c598486439d4d9f8b447bdce1e7f0a2c19de525d38b55211cb333c7609d7 +file antiref92/train/shard-00160.parquet 127748688 b1f7dc8aaa56280b06cf14c1eab2e35f29ee32b1c4a7ca8f3b44af89c3c97bbe +file antiref92/train/shard-00161.parquet 104982617 9375a82c8d5da2911317ad09839781450483d86797e97c0c94c0b040abaaaf05 +file antiref94/train/shard-00000.parquet 123952273 42f0992b08b1d4afc2a6a05df945a67afe32bae4de0f576523cc0b6dac0b6a80 +file antiref94/train/shard-00001.parquet 123400276 56ef6d7a89d06b429414c5c3a0a224753316b258cd2f855e1f12f95741218357 +file antiref94/train/shard-00002.parquet 123194874 686ed26d3023ef89be77d2a727cb181d7e830c47370a233813999369b2481d09 +file antiref94/train/shard-00003.parquet 123738209 65b7536611bb391db04e809907b147adb984fdc7bedd8dba75d6d0ad86be038c +file antiref94/train/shard-00004.parquet 123450083 deaa6c66b980b0c1732fd49eb282c5d1356e684b1e1ae41caf1419e14c0f3c14 +file antiref94/train/shard-00005.parquet 123569853 2212fb364a878f98ca0a967cd9a7c90e37f1a027205e5eff19345810adb1405f +file antiref94/train/shard-00006.parquet 123173684 618401183d3a7bf3212c0d7f10c2d680691a938ccb29c35cb0ab7ca11a29097b +file antiref94/train/shard-00007.parquet 122989486 0705fc8b93ad4ab475910e8e0cd96e1443e6512b2324a796ad5676ac1ffc22f0 +file antiref94/train/shard-00008.parquet 123432385 23c3cc9b7a0f558ae1186018458edd09e3f21dd7a480fa4d03524bb3fc169c22 +file antiref94/train/shard-00009.parquet 123080394 9707f108070b029a740c3c5509ed7d18fa8fb2dbdecb9b5970e23a40258d3417 +file antiref94/train/shard-00010.parquet 123575631 56190999661cd54800cf883bad4be3f99815dffaad8f2b7a72801823886068c6 +file antiref94/train/shard-00011.parquet 123174409 29937308dab6328b0f8db8d6ab13ea377462b287ca82158a002b25b306cca76c +file antiref94/train/shard-00012.parquet 122831867 df2a51cd7758c88fcc1dfa8ee807c70acb273b5772397aed6533a7765033f9ae +file antiref94/train/shard-00013.parquet 123402649 c1aa4cdfa9c54d4524a2a05e97d8365563b3d943a87478ecfd6924a7013850c0 +file antiref94/train/shard-00014.parquet 123885530 33160c40c85c1d2744f171c34b3332c003fed76a22ed1897c70cb7c770f9c021 +file antiref94/train/shard-00015.parquet 122889972 a5e736f2853f1ab049dc801c6dd373aa07bc2c9d9abe642180018d010e273178 +file antiref94/train/shard-00016.parquet 123736980 a22644fb2fa9b51f0dbd7437dff524379564a1535878f9d8a68c9d6ebcc26e19 +file antiref94/train/shard-00017.parquet 123749103 8b2958b98dd40c163af89500a861cb9d55ecc017dfec0c06a8860d7ee40cceda +file antiref94/train/shard-00018.parquet 122573831 ec1df02c94063a00b58eb802e38f493183cb5975ced1c5c535d5bb154944ff74 +file antiref94/train/shard-00019.parquet 123775665 8e3f14839538217178650b922766f2261d7fd2db12fe12cc801df3acbe9feec5 +file antiref94/train/shard-00020.parquet 123562190 c7ea8746d38ba90d93aeabf28ed8917f7912ceab6b326dac628f52a3b6b85871 +file antiref94/train/shard-00021.parquet 123319085 c7efe790cbb296318fdc7e5b7ebcbc2e4483e75d9e0c1b5c397690829d99d690 +file antiref94/train/shard-00022.parquet 123069530 6c9f48e6a9856972def9ee9854196881a32394a58ed8277000ea0278d4f20f29 +file antiref94/train/shard-00023.parquet 123415056 f7bbd1c62a35c3c834f91d5cd9b517719974788a922f1906e9be8d1ba41713ed +file antiref94/train/shard-00024.parquet 123519548 d7a9a0bf9efa79127e06bfd9e7e3cb1c01cc1a2b9ed50f2cd61c6fd0659ed90b +file antiref94/train/shard-00025.parquet 124038446 9b5f37575ceb1335e6a9bd758c0e2f6ecd61f3f459a090cedfb2d22cf5c972f0 +file antiref94/train/shard-00026.parquet 123297357 570e04b24cfb8d0a1f2445c6ce218f9d40ea34ee316191679fbc6ef2c5b972ff +file antiref94/train/shard-00027.parquet 123858108 90237494600566c2bcb0ce075b608b830b604ba3a97a6f0f25f2a5220b095496 +file antiref94/train/shard-00028.parquet 123219652 fef7fd58cfdbfafa5a7e9daac18cd2cafa499ca63b577303e0a6fd197278d416 +file antiref94/train/shard-00029.parquet 123646077 801befc6fbf94aa76e5242918385228c1b4ff94e29cb0916cf3673d231663159 +file antiref94/train/shard-00030.parquet 123323651 3772f24b5a15109a132bfefbe7b4dfd5269a510817192ec0c6818adc41695b7d +file antiref94/train/shard-00031.parquet 123963129 2e22ad7e0b4d03851bae2c38a419adb87775b5e07cc22cedddc9d84bbad5205c +file antiref94/train/shard-00032.parquet 123934945 72f76d5a4492a97bb6de6a27e01e6fef673c816c4071e3ef2ac72c8c344a8512 +file antiref94/train/shard-00033.parquet 123613588 9cb1b73c05835d0e9a3d6272470437e530c6cdb471aabcca29fd25bfb38b25d5 +file antiref94/train/shard-00034.parquet 123439345 ca029b81ae2f7fc3039b2b72e3f12441456e7bfd9063213475f63f287a359aa6 +file antiref94/train/shard-00035.parquet 123475441 ebb8f6105b5ee7f087dcea09ebcd7f4e7b7d65adc1bfebe871903b328dcc73a3 +file antiref94/train/shard-00036.parquet 123385608 9afb926a17217a28a555adc7b37f72c7760bc73c8610457d8625d37e24fcf725 +file antiref94/train/shard-00037.parquet 122803736 b626536ee26af3a9319accad76dce032af352f5f0846d457f6ef00a5c6d0ad29 +file antiref94/train/shard-00038.parquet 124364779 769519b475663ca5a39fde1cae842296805503aaf5a13a2666a25a45b048a8f1 +file antiref94/train/shard-00039.parquet 123811346 a7cd38b572cbd4920e4d30764760e4fc0c8b643be5378aa6916c9e7d48b09642 +file antiref94/train/shard-00040.parquet 122582432 60c25163bb7a9d42e4f28173b92b2f158c6b033d92d2a1d75021140114f77ff7 +file antiref94/train/shard-00041.parquet 122773999 ab369220e78f66ca87674ab055374200b19377c81bea1baaa6e22203f67134fe +file antiref94/train/shard-00042.parquet 123455444 0dca31644862e4bca7690e6c42fad105ed8c3003fe0b22b35af3961b11dbf71c +file antiref94/train/shard-00043.parquet 123154895 f4252d60dc904f30940ca0395e13bfd68462d3173f136ced564d88c1b8917d7c +file antiref94/train/shard-00044.parquet 125223939 9bf28d167bbf1b3cdb5cabc54ee601314759bac4e37d1f0c7077d7019fa063cd +file antiref94/train/shard-00045.parquet 124218361 cabe9adfa7a60e8c383d64dabdda8db7040e3d5878a59c52dcc03c95389b8ba8 +file antiref94/train/shard-00046.parquet 123132668 95554f2612f43b9c7211d8b9cb549f83416f5348b68ed18c7d130805d6afff10 +file antiref94/train/shard-00047.parquet 123614385 956aa8ef7d14eb994fa0d84c9f055b73a15bfc29bc9e24c8814dea804c845510 +file antiref94/train/shard-00048.parquet 123857939 57f988e97493ad3f8ba8616a55a3c00781ec68db59acfc8482a18c3edd1c5d0e +file antiref94/train/shard-00049.parquet 123654082 bffd8bb206a6d68995e888da5ee3cdc5978a2cd1ac34a41f7645012e7c57e7ae +file antiref94/train/shard-00050.parquet 123064836 125f8177c39affeec19321ce68ff49b78469fd1e7262fd7cb48cac1b61f81f52 +file antiref94/train/shard-00051.parquet 123265213 87d1639a680af068716a715dfbd1920f9a6a8ee3d2f2101e2f52ef3081f6bdc8 +file antiref94/train/shard-00052.parquet 123766128 0bb8e16b6e37b369145792a530790dcc920d0aa68906ea8bc2e410d9621eda2d +file antiref94/train/shard-00053.parquet 124181913 6334e56a98421189755b0f42690e003ecdd6c49db896c3f4d9ae44cd8b2ffb96 +file antiref94/train/shard-00054.parquet 123842811 180bf1767bd6011b2f1cd88bf6b389c23dda12897cf78a84d1030b493e064a04 +file antiref94/train/shard-00055.parquet 123894934 b72313cbd8082e5aeecf8f58ade1bc7c5c879037104c8dfff75b6c81c656bce4 +file antiref94/train/shard-00056.parquet 123303779 569f839a666edbe75cf6a41cfeb17ed714be9e46e817410029b2ee6676f626f9 +file antiref94/train/shard-00057.parquet 123363128 e7800959304b0acf61af422672850f03fafb69decbb806e9034a4bf4867921c4 +file antiref94/train/shard-00058.parquet 123725133 d1f529967284e3d5247bd432a0e52a3dfa0d3b4e8df363b50da9cb078f44dced +file antiref94/train/shard-00059.parquet 124081968 c729b597bd10b7b38e680bd6f778d0d502b6a87593e0aabf4fef78865a5affb9 +file antiref94/train/shard-00060.parquet 123431203 11a2e43be9145034d740c36cd647b4a91e09de17fca0cded85aea605d0adc0f9 +file antiref94/train/shard-00061.parquet 123062490 8aa06b4d535229c420184a60f906960532c27e09e804fad58375fbe50b8e4f57 +file antiref94/train/shard-00062.parquet 123602179 d7c0eaf4b953ba2956a994756505ec5c22ce67e0142936f3dedb985b57c7e91d +file antiref94/train/shard-00063.parquet 123967877 8f991df095c44ef31b7ed8a4c65297ed0cde949ee1c115de18c40dd5468e80ad +file antiref94/train/shard-00064.parquet 123644862 302da4f0764b87e91d5a377637164e3d2c8681ff845ac33b6181d5545984a833 +file antiref94/train/shard-00065.parquet 124010237 5c44faa79bfe3d703f456ac1707a70cdee5224df433197f33892f204d778a2a1 +file antiref94/train/shard-00066.parquet 123678161 d66ca5012085bdac0a4f381d7b200d70dcc63d98b65dc62d2f7b6ba75cc65c6c +file antiref94/train/shard-00067.parquet 123512430 83d3fc847e97957ce8d2163a2bac9cf75a2d667720fbfd7e4c3822efd41fe6fa +file antiref94/train/shard-00068.parquet 123205169 d5453955855871dc9e2909d7363f846c5f101769dfaf2c9b7a1f7eea9a4a2a03 +file antiref94/train/shard-00069.parquet 123243711 9e3a3a1ce705a5e484692a26aeca79bdfb4239193e1769824b89549327d98dbb +file antiref94/train/shard-00070.parquet 122825989 82073c6a64981d4a0dc8f8e0f57d715e164c1bcacb24ab21d198060ab0704695 +file antiref94/train/shard-00071.parquet 124328922 31db73e915005c62a80a38ceb8807197e606f824ac5c2f4a5f6a4b904810727e +file antiref94/train/shard-00072.parquet 123811447 6b8d58a6389632a237c4adacd1c14eec770eb5a239cb377cbbd17c1a5830e545 +file antiref94/train/shard-00073.parquet 123013646 6d4b50a6d692d8cd37ecdc81a5480440343fa57a870977b832bc0ed339af4c8c +file antiref94/train/shard-00074.parquet 123827161 866235b9209c6a5987dea9b460a07f31f038f06ac836a2cd4860f631257b14a0 +file antiref94/train/shard-00075.parquet 123562764 59e23874e976c01969ac4627b9790ff2d2e1bf188175c9c70407b0117a3b215f +file antiref94/train/shard-00076.parquet 122942391 3468f67b0c59f00ef4347699b8ad93dc4d9da57c74c6f138549514411c680435 +file antiref94/train/shard-00077.parquet 124432599 3f2421c55e819e067726e1c63456aa50a7162633f423fb0b249f46f2d5333d29 +file antiref94/train/shard-00078.parquet 124061976 da6fa88c9c728ad21453908a939fd2ba7bdba60e9e80e96df9284241a2c2e2a7 +file antiref94/train/shard-00079.parquet 123015002 072010d706c5439dba66ab73cd3dd38d07b071788138a98549eaa1d9776b388c +file antiref94/train/shard-00080.parquet 123852917 47e7325f8d620aff3f63b52f7cc13bb11a07262335b7206ecd6e114fb10ad157 +file antiref94/train/shard-00081.parquet 123332497 52c76c79849e847f95d69d403bd50b3fec2b7e7189f76238b200a525e93269ed +file antiref94/train/shard-00082.parquet 123323594 b06ab1bfac36149619cbc5481c1e3cc7117c2ea34ac2926e334aecaf207cbc0e +file antiref94/train/shard-00083.parquet 123653870 7d4c04684fda524bb26b035ea8e043a03b382d2a8ad824f82c215751beaeb880 +file antiref94/train/shard-00084.parquet 123700285 1436dadfe6ef2ee641691aa04926e4311775de54009bd441aba5d7a8de3f1afa +file antiref94/train/shard-00085.parquet 124423263 4df8f7b0304167028bfe07ee8bfde415b33ee14638175bcb3b71b9c790fc4251 +file antiref94/train/shard-00086.parquet 123638034 df6ef93fb1ea961f012de185cc26585e01181f1432ac2d29eb298f0090cddeca +file antiref94/train/shard-00087.parquet 122688106 86e386134c10ccdf1640509af1c480b8bfb282f6644574e2a0d530b065555482 +file antiref94/train/shard-00088.parquet 124184868 a74ae9a98eb91e7026c8b33eb67f0bbd194a2015795aa7abce9cfa943d8d8ff3 +file antiref94/train/shard-00089.parquet 124175175 f76906fa110f23754420e7069320fa3656bcc1ed8bf4d0b1eefdfcdde0a1ac8b +file antiref94/train/shard-00090.parquet 122909173 0854320ab19c1b45e7a14c34bffe10725cb457e7c5f2eb2289f60b6b2a4403c4 +file antiref94/train/shard-00091.parquet 124737869 cb865436ca537de4ff4706384328223c7af4b766c458fc263d3ecab9b5da2715 +file antiref94/train/shard-00092.parquet 123726868 6bd7d4782a122ba43cb258af6267bb503ac3afe7aa89a59096ec17bf36b146ea +file antiref94/train/shard-00093.parquet 123305620 9e88fedcb30d0f387046e1ad4c5cc5dfb6eed0361e2e165a1b8aa97b3647c898 +file antiref94/train/shard-00094.parquet 123953937 0defd502233ab2aa667f0ef74af1a236d2b6248dd4b24947cece6c4f82ec32c7 +file antiref94/train/shard-00095.parquet 124287601 6bec81ab0cb59a02a6e4c143d019af7a56950b5345c94894fe1bfb58d25525d3 +file antiref94/train/shard-00096.parquet 123828948 6769eac763efd68227b5040a673efd908ea8cb9ef7f7711fbce25aa5c1b0557c +file antiref94/train/shard-00097.parquet 123133219 6d68034953fbf8f3ed2ed0c8e5965ca253c61d3512cddcee66dc40167be30830 +file antiref94/train/shard-00098.parquet 123090574 9fe719f3a3eb4818c8bd2b2b32408773d83d869ea594cb2b1d0ae7892bc7832c +file antiref94/train/shard-00099.parquet 122869980 cf9c0170d64220fff92db463c89eaa7a8adbde554f5ce71c237e11e66736b571 +file antiref94/train/shard-00100.parquet 122899501 90b7ee25fd82c536611efbb3c69d5ad7c654e9db88d650c13f57f90d2f1047c0 +file antiref94/train/shard-00101.parquet 123854557 ad71228ae5ca492649565c221fb73bb19839f474e5f7be840862891b748d5a04 +file antiref94/train/shard-00102.parquet 123709927 77b99f319636111bd22b353810e178bdc6d4a9b072c765632887835ff9ed25e7 +file antiref94/train/shard-00103.parquet 123713679 e1c82814fe457ea71e7084a317001dce2cb3cba9c6d4e202ad1857ac3f95e508 +file antiref94/train/shard-00104.parquet 123171946 2d4c123d0467bb393d75fab655240f48fe24d633f66cd05e34721f8d0fdd1ad4 +file antiref94/train/shard-00105.parquet 123657964 59501aa0ec90315391a3a316b52b0d02107e6f3fcbb07cf5fbf903a8a4db0a1d +file antiref94/train/shard-00106.parquet 123459571 d526e6538d1b017f1e381d5f41fa4f08d324a1cb13d40dea7dc113257b51f26c +file antiref94/train/shard-00107.parquet 123822170 8b15d89bd4a238effc6569ca609b77222ff3644e8dd05002bdee23a2988e931b +file antiref94/train/shard-00108.parquet 122676879 915838e4582a248a0086d2c2023d44758cdaad2c00e5335c31a9412093e50db1 +file antiref94/train/shard-00109.parquet 123977100 012df444baac3ea981fb90b6a13962aa28a1f09da8b51c9e09c99a7e839bd4e0 +file antiref94/train/shard-00110.parquet 123814774 e1ab38835d77d86b24d01c77123bf0d6f5f39af6a30bd2fd6e4d348ed7ca79e9 +file antiref94/train/shard-00111.parquet 122562022 c7219a852017da0b36c5d332e7536723d5a52fbd4b5a6310b15fe099cba297e8 +file antiref94/train/shard-00112.parquet 123511896 d6a56e3d43a5e4e1fb97b5ae80d81bdd6b498e3c89a39a96219280896c1242ee +file antiref94/train/shard-00113.parquet 123257292 a21f4e97dbba51bfc54cbc7d84f539327ec316d86360dda0451ae5440083f723 +file antiref94/train/shard-00114.parquet 123706090 73f4559d93827844692c292f1f8b034fd2192039a6d117c9e3fc5d70e4bfe82f +file antiref94/train/shard-00115.parquet 123388141 f91af4f50068c4840afc64d76e13ace36bbb1107fa257baca6a2c47ac425b4e8 +file antiref94/train/shard-00116.parquet 123880240 1edfe2d28124728a3366dada528b1b22a21f8f332291225485ae6b5aaac2db27 +file antiref94/train/shard-00117.parquet 123989295 fd7682a08f80481332d24cbfc46dd530aeb5b19a0bb8c5b36751c6a5817a5b3a +file antiref94/train/shard-00118.parquet 123412936 f232aa21c9f080076cb66086f002cb70f30685a49f8ab7150a7bf55807b7c34f +file antiref94/train/shard-00119.parquet 123121604 27da467358bfd35df5382037ea6c27574f8ce150dca67396bedd0a6037792f80 +file antiref94/train/shard-00120.parquet 123755924 5d2d6d1869f59c7a1e0ac3ad75392d39489a578e4366c97f30d128ef79fbe5e0 +file antiref94/train/shard-00121.parquet 122711552 4a06537134b827c9fe2f958926f58a2a133786b157033ed995367797ee300985 +file antiref94/train/shard-00122.parquet 123358643 6862899130fb6eb9103ece1c20a249588cc08f4f266d6f994ddfa47c8de71a62 +file antiref94/train/shard-00123.parquet 123570592 6b04a14b0ed6983c9ce7d679aec9c656f01d8b4976a2ab7261ec149148429f97 +file antiref94/train/shard-00124.parquet 124736205 71c1b8fa357db914faafb796a7563ce48c76b2ff4d1cca100b74b3d48d9cf953 +file antiref94/train/shard-00125.parquet 123792924 3605c2f781acfa818c05db860cb9d5eab9b40c1311e62a422719591297b2c32d +file antiref94/train/shard-00126.parquet 123769580 61ca4acdd0f4702a85fad6eeb83ee0679f2cb6a1445714f9d55f3689d324aeb8 +file antiref94/train/shard-00127.parquet 123369481 d87ab7f4daf791dbcee97cdb29bd636752dc81dd120056893a56830d5d99824d +file antiref94/train/shard-00128.parquet 123256308 541c2589bebd1f80288d9eb539f559050b6a062feb4d35641d258492ac66b996 +file antiref94/train/shard-00129.parquet 124251653 4fcbb3222bc0b3d50be03c9008dc262e05806503dabdf449096cc85e1a846fc1 +file antiref94/train/shard-00130.parquet 123456420 6559c21e3d1acddef7143f509f060edc945eae5fe2f4dc1379f0741a49f6532c +file antiref94/train/shard-00131.parquet 123869073 c92098dbcfdff171509ffd4d38503cd5415d4ec96970e3ca48b59a8f52dfe5fd +file antiref94/train/shard-00132.parquet 122946672 de5af587ad08a31e2db074ec9ee366b3e9ba196790058dfc5ad3afa1d93696b2 +file antiref94/train/shard-00133.parquet 122614171 0a22ee373f9fb35d06ad07b578f5d05b96e6ce967ceb344e889430287b582701 +file antiref94/train/shard-00134.parquet 123141500 4a743aa0d3f1317915ec0b18579cf5518ded5dc19889674064c50fe32ffc6592 +file antiref94/train/shard-00135.parquet 123459055 4a7f9b0d09c28cc68d7f7e1db72f2a05140f0d6dddaca557d206d9e80875d8b3 +file antiref94/train/shard-00136.parquet 123453178 81d5cdd6bd06e30b487e8bd29d8c31ea945319cbaac1a9fe9c67c3cda2fb91fe +file antiref94/train/shard-00137.parquet 122609275 3ad2b858263fdd72bf85ef45125c805218b1ea21ddb30d2121427390f11295c8 +file antiref94/train/shard-00138.parquet 124074164 1da314cf3bbdaacea21636b32a63a18d26e2ce480d640535d52f57277364de11 +file antiref94/train/shard-00139.parquet 123894638 eaa8ded029e0c9f787f5ece9a7a2ba763cf335a05e9e81b78e3c77801c248f35 +file antiref94/train/shard-00140.parquet 123071390 25535203fda65c711a123944190bc7dae1d922507897091b496df6b3250d231e +file antiref94/train/shard-00141.parquet 123564116 06b79243dd911230728becb5203db62823f1c9928fdfe3a34eb8a9510a6bf178 +file antiref94/train/shard-00142.parquet 123951104 5832fc37ac3d3d11259a4b4874d17b799f5c08533a695289e8917c92542f71ad +file antiref94/train/shard-00143.parquet 123388713 fdb88ddf0c4639a5ee54d080a0b0820e073f421486d317a9ef8917ff323d59cb +file antiref94/train/shard-00144.parquet 123761068 4e38df1606ad44eabbe18a63ba99bcc3c9c260d49400e7eb9a758dc7319905f6 +file antiref94/train/shard-00145.parquet 123491027 b252e3019b932b6ac178190e29f9ed8e5ebd5897d209d6bf46641b887cb36fb5 +file antiref94/train/shard-00146.parquet 123330996 a0fb6a6e78c165c22ea5e611e5c1cb19592f641316ed938caac3ca2a9e45c558 +file antiref94/train/shard-00147.parquet 122959526 8fa94c7b8615494d2a13577cc0ac424edc8731566c1f8442e963da7a8a752036 +file antiref94/train/shard-00148.parquet 123331881 7e956ce0e461a0a1d9e771c6d337750d67f8971f99f57dfe7544cc0c0b51b702 +file antiref94/train/shard-00149.parquet 124511226 e03c31f475c807e0b1148e7708e68bd12ddf226a18a4d0dfa261cb83cc762212 +file antiref94/train/shard-00150.parquet 123932067 609881c88614f6ff8be7b87d0fc1604b630c17efc78ace0c22710bd24beeb162 +file antiref94/train/shard-00151.parquet 124195342 ec21916e15c06bd2f8eef3a9a57c244d3330c88828c62c7cd7fa9f6aaeffd996 +file antiref94/train/shard-00152.parquet 124039158 caecfc6558de866a70aa0070664ada7920c82a98e349a1ded7370b7ee8732ec0 +file antiref94/train/shard-00153.parquet 124141247 fa8c42ecddbb186c3280205fdc4090b4a82b82f6e1508329ec75988c1dbd8443 +file antiref94/train/shard-00154.parquet 122885969 5905dbe22a9351aebbbf6f5166a5ce92bca928cb30d6b52668273cc07279a5ff +file antiref94/train/shard-00155.parquet 123210838 b57c397f5be8ca23a22defa6e9fed8acd2980e39c308427218b255e327349d59 +file antiref94/train/shard-00156.parquet 124375882 ff2c41b5fae0b95499e08f8279811a3f0c47cb793aa747a0dcd558c9e6e85180 +file antiref94/train/shard-00157.parquet 123555286 2626ef5bf1c012874cfa854cf5bc1e46eba47417c86589cae4cccd93d5ade984 +file antiref94/train/shard-00158.parquet 123825858 d2620dd1b458ae078af56828afab576831f7b3cddefca1e562755d06913ac26c +file antiref94/train/shard-00159.parquet 123321296 52710ccb32df1fb5c341ef639923ac859f692a44d52ef5d0d9981a108103cce4 +file antiref94/train/shard-00160.parquet 123578497 513852d886686c825d310b7e042704bd2301e542e3f5a8616e98dc537b51a7ef +file antiref94/train/shard-00161.parquet 122987782 74923dca2b794ba48dc7935477082eedcde9d5d6fd942fad6a6b2d5ef4ac33cf +file antiref94/train/shard-00162.parquet 123733611 3382e434a62013bcc4bfa85db33181eb839e4ab36be4d627db8f47a37e68ebd2 +file antiref94/train/shard-00163.parquet 123790821 4967bb5cd9015e073856c8422c03dddef8f7f2c2d7f6deecdb7ccfbc3a3540fb +file antiref94/train/shard-00164.parquet 123575317 9722844c2a2f4b2f8bda2a0bc032ec3d9aa516ef99edf1c8387c70b62a577dd2 +file antiref94/train/shard-00165.parquet 123683079 be0313087325a197eeb28c49a2da57d50729d5a0f3e981ad5382ce4be0c2d3ef +file antiref94/train/shard-00166.parquet 123088728 7509f33ae293f84510294c9ef216df588ae25078ed6af224985012d8559d87d3 +file antiref94/train/shard-00167.parquet 123672466 ee89c1c5966526f3ead3b76a1ede0b1ea8a48049a7bdb8e00a12f7cbdb15d334 +file antiref94/train/shard-00168.parquet 124075799 fc2d0255f141d2a9e552bbba1cdf19b94db7d9631c0b59f477d0ccf942ae6771 +file antiref94/train/shard-00169.parquet 123308557 d8170b23827a321b57d62f49a23abc64e0016c9a71bc59b69bf04224e049a400 +file antiref94/train/shard-00170.parquet 123943797 2247fd4038be3d38fc3a10f297bc2433b6ec9dcfab1230d90e453e1c64a2df69 +file antiref94/train/shard-00171.parquet 123625540 763236b6bd1a75aec4be425200b29458360de9633d5597c2b46de7454a4aeede +file antiref94/train/shard-00172.parquet 123348045 c87e84939d68303eea88d09fdea28fbe6950c8d798a57ea63615345fb4130bad +file antiref94/train/shard-00173.parquet 124428885 e4dd58602278a7ba0f7511596282e4065bcf05c531102c1708a3c3c6f2b1b5b1 +file antiref94/train/shard-00174.parquet 123186741 c4522fef17b0d1f3d519e21d3afb1d21cbfaa4f63845a4a051c7e1813d481ce3 +file antiref94/train/shard-00175.parquet 10280395 5a3fb3fdb0e92f14d969038a4d3bdef9fe733d2f970f852e4fbd26d46b408210 +file antiref96/train/shard-00000.parquet 118642154 5004eda250abbcdad999c8e7c0b71b5bf2275702cd34d8aab727be09cffb4b37 +file antiref96/train/shard-00001.parquet 118548201 56ea0a7a62871771047fa2cca4fa500dc28e32c05bce9211b367a1b5b4bbbaa3 +file antiref96/train/shard-00002.parquet 118677030 6603f1aefba90b39bd9d076d8eeca28251de385a5aa748bad3e37364f0acdaf3 +file antiref96/train/shard-00003.parquet 118340313 dc86ab956111cca8faf44dd6746c23c6ae37e9772074c1e71117951c8e480639 +file antiref96/train/shard-00004.parquet 118583068 8044110a767654742fa959fe6e757b0895f206aed1cedb606a0e30fef86fb474 +file antiref96/train/shard-00005.parquet 118705343 84d935a370a8336fc4b8f4b5b087daa164dabec8e1693d8b6e92e614c75a55cb +file antiref96/train/shard-00006.parquet 118664121 20950574e0ed94e16de7e8fbed7c9323d1d431208c97b6c98def68dbcca83526 +file antiref96/train/shard-00007.parquet 118248110 01e85ea24f2b2f26a3a011831c2b9eda5ef5a625d9340e86ae353032cfb27441 +file antiref96/train/shard-00008.parquet 118304042 a8b322ea0408ab1dbf1e22014babfb484ed477136dd8139bbdda09bc7e5f7864 +file antiref96/train/shard-00009.parquet 118668228 f284aacd0ddb360f3d47d3e1e231f6ae45f5427881712d4464e5b288f22ba8bb +file antiref96/train/shard-00010.parquet 117990145 9a5919006bc9146ed62bf62d7d3aa9a8d3a9a5051cdb83c49ca3a8096fc090a1 +file antiref96/train/shard-00011.parquet 118259076 b3053fb5709de07e1e2ac79a1e3e5e1300263c6a7ae7375430468316eb300839 +file antiref96/train/shard-00012.parquet 118694860 9ef1f5e2a5c063d7f5213b65b8269971c22651b0f352cc3af1ffe4e9eb7f4509 +file antiref96/train/shard-00013.parquet 118192046 01fa93885950f95528746f9263b3478b3fe94970d6fc921166c6eff00305428b +file antiref96/train/shard-00014.parquet 118005645 ea506efd9ac9941d26720d429b1f7413e2200111418a71c75a340dab6b6d3e4f +file antiref96/train/shard-00015.parquet 118605006 e844b94e7b66839309935ded011a52a1e7abebefd0bb77cb3ad274b297fd9eee +file antiref96/train/shard-00016.parquet 118719581 f32d4758072333e9175c1d5222b7eec3e993531868bf540cf24e33966c249783 +file antiref96/train/shard-00017.parquet 118354926 a1ee8ac581d0d29d7102218b1c8b72980588912a0eedcbab63eac2580510eddc +file antiref96/train/shard-00018.parquet 118369224 69b7370ae2d06081ab5791ed6286e63e212f4841b5bc22896a6ff908f17951cb +file antiref96/train/shard-00019.parquet 118632576 160ca5c422056675c09c48ce3991cc328ccee0c59c699d9e50cb5b40c130879e +file antiref96/train/shard-00020.parquet 118627499 9915b1e39766ab0abd9f7260329931e0dcb7e72a26fe58bbca09000b4edbf276 +file antiref96/train/shard-00021.parquet 118279345 82e4c935e46eed087dfde1aaeec8ad00b4f8cf24fdac08926dc4d3cfdb04265c +file antiref96/train/shard-00022.parquet 118735948 5bf066bb460500e1d431ea8758fcc98f53e570938dc2c65559aa4f2ef3b86771 +file antiref96/train/shard-00023.parquet 118641409 85fbd08f6f00ace28413a9c5dbcb1f2ef4aae79c436c614ae3fa071b8c5d2cbf +file antiref96/train/shard-00024.parquet 118495193 f3fdcd123a8e92858b5c6bb28d0b27dcf3ed768646fc9c219faae98f544d460d +file antiref96/train/shard-00025.parquet 118476261 5ae7a202b982c3218ac6ac69746ea5634adfa5b1e1491e145d1c9520dd76bc34 +file antiref96/train/shard-00026.parquet 118229959 e7eae6806e3b4506d1f62a495109f6a6c81ae78a744b2ff73907f85f812f8bc0 +file antiref96/train/shard-00027.parquet 118614595 27beed8450d9092b9bab8db2fbf5bf7807fb8b8d47351ad24c76ea1211b95dbf +file antiref96/train/shard-00028.parquet 118637734 1475b4639fe5604b2f60b20fa433071936a9616ee150a1d1fbbeea7fed39ae91 +file antiref96/train/shard-00029.parquet 119023250 9d11de8d55ef3a4908305a23e3cf41b3b0de4cd86d395244d6349c144e9eaad2 +file antiref96/train/shard-00030.parquet 118186975 d071717598c9587e436d8653a47b5c6873205ca9b837c6b8f62916ab1d2821e1 +file antiref96/train/shard-00031.parquet 119153295 b5ce6cd2761c2a2bd6fc16084d2c3136da49fa57ccf84225c55285c97dbc9f30 +file antiref96/train/shard-00032.parquet 118451667 4b6abaf1e963ce20d13cc4d8b23f6f29bd2c66daf8a8e80cc3ff856afab35f6f +file antiref96/train/shard-00033.parquet 118464059 009b05118e012db35d3d2a6f224aa6c42f2c45be451dc5b5b5a428ca774c3ca2 +file antiref96/train/shard-00034.parquet 118556648 39b6d3be8f8f6d5efedee336182cc819ae4ea3dbbbab7a3938a2bf9d654d16c9 +file antiref96/train/shard-00035.parquet 118359302 a73afacca7894ecdbb6cc04dd05746d3ac18020461c01df86ceb9bf2ae70d2d0 +file antiref96/train/shard-00036.parquet 118916526 b8e4af72c3483a21384416e076a9ccb3178ce7a91b11466309284a22e799fc17 +file antiref96/train/shard-00037.parquet 118316214 59266a25bc5a94e98160e9ca72a9703f96551a72b5c4c68391a100ab41f71203 +file antiref96/train/shard-00038.parquet 118583679 512218fb550c757c0ee3e3570d31f7c3a662a6026a9c638fc8ddecb157840cbb +file antiref96/train/shard-00039.parquet 118321590 8169ae4e77796fd54d574dd5b3b61652d80337cb658506e40a6feeb3f297472e +file antiref96/train/shard-00040.parquet 118609908 9e8a1ddd3d871024abec9d4da6a75f6342f1367bf8d9a085b937ea4d5ebb9c88 +file antiref96/train/shard-00041.parquet 118340233 010fe6ede5e5748b9b449ea5209e0f4b3cc690df8f3d1872ea1aa2386e1038b9 +file antiref96/train/shard-00042.parquet 118214629 553368768e7363e816bca5e7f03351bc0943fb0a01b51c774a11d2d28c92ea0b +file antiref96/train/shard-00043.parquet 118749657 712538c923383903db659addb0cd6d2b0158001772a87052813eea15dd1960a5 +file antiref96/train/shard-00044.parquet 118560286 a2a1fbe115d24808fa7994505cc90612af628de268e4a3b7344764e780c301f5 +file antiref96/train/shard-00045.parquet 118569272 0345cf4c3a3f917a7f47d517dc53e8bd41df702f252dab45dcecf0955b758e11 +file antiref96/train/shard-00046.parquet 117972979 2a5f9c46effce81d4538776244db5e039cfc278934f12297f75c62441e778485 +file antiref96/train/shard-00047.parquet 118477408 abb961f6cae33b04563064b3cac767e9faf811cc93a0936b83c517e27b9a2032 +file antiref96/train/shard-00048.parquet 118551258 9d26d1d88bfa56307da8230365d78912a5b49f41735cbe5d2e36f401d9e4c2ab +file antiref96/train/shard-00049.parquet 118173607 6564075580bf6d647b6b8e2405098ef9ee233d0c84647501f1b451181903c729 +file antiref96/train/shard-00050.parquet 119524951 23187cea9fd0b6a4157a87de2fcdffef5153e36801f839860fe6b86e569147cc +file antiref96/train/shard-00051.parquet 118786941 a24fb11d61eb20de779714608362f7d398c16370dc95996a792648e424bb8e86 +file antiref96/train/shard-00052.parquet 118628605 20f6ec1b04ee68ea51e7cf53e3181087fccfe8b98b8331454ce0cc9460d02afd +file antiref96/train/shard-00053.parquet 118600589 9ec858ef89e49580fcf274d72f3aa4f1cfebf1d883895a1dd05c61d5adc43b48 +file antiref96/train/shard-00054.parquet 118533004 a96a0026af09f7663ed9dd339d4e114cd7489c3bbdaaa8ef57d89ed5cc070fe2 +file antiref96/train/shard-00055.parquet 118531870 5cab0222c074214a231b9fdc7cd11300772e5f0adfadb07097dd7f9ebbb2c48d +file antiref96/train/shard-00056.parquet 118525515 635dc199d8c3cc45ef558cffffcb05b77d871eebcab6085fdefd9429f2ab3d97 +file antiref96/train/shard-00057.parquet 118244268 b734ca970e3ddaf7eb8bc796cf5f765c84cdab585ff608710089757435a495e8 +file antiref96/train/shard-00058.parquet 118429880 ff8ef36a837e6bdb1ff88193d6c5ea59afb8e478c4ff406740fb66c9abc6e401 +file antiref96/train/shard-00059.parquet 118744082 698f96a6047b50d6eca9c9f9df30e4ae8bc84dd4d05c9521b904548150efe782 +file antiref96/train/shard-00060.parquet 118773154 ead448139221f8638e0d15b4b5721463201498f0b8355c1fb6badf8c8027ad5b +file antiref96/train/shard-00061.parquet 118714149 0dc4a0d11f5caea33c8154233d413cebb66aaa3a4d92cdb237e24ea36fcbc92a +file antiref96/train/shard-00062.parquet 118390164 61d80f6bef205d3213a8a1886c8cbe0b7d4831293f79e62a3cb603676a86ad58 +file antiref96/train/shard-00063.parquet 118786775 6ed94ea6c0e7dc169155335640f6ef038243ff2b112375b826044c7282da9954 +file antiref96/train/shard-00064.parquet 118680596 2daa3d74dffc27aad56a43393352b94bcaf13b0134838c562d907d16a19f5867 +file antiref96/train/shard-00065.parquet 118215280 75490e9feff0dc6e70f7f6a0047541d2e10a63e95644cc9a2cfc5f3b7b0ef0de +file antiref96/train/shard-00066.parquet 118511503 58c7f02b50d8c161ab81a3692d4dd1deb8ec2cafb5a43788c3edc1d711ef2d94 +file antiref96/train/shard-00067.parquet 118834365 00c15d1414293ec4bb37201849237308de5d64d68e6eb67c6ef8bbfd05423247 +file antiref96/train/shard-00068.parquet 118699514 7ba2eb5d1cbbc28030a21f647d88ccda01ef9727c252019db4396acb8f86f331 +file antiref96/train/shard-00069.parquet 118186965 9ca4126feaf4149a90e6ad1740af5c2afa193cc80bf38134d0107aeb9d47b626 +file antiref96/train/shard-00070.parquet 118472295 932c2c6bb0e9e1910914ad940dc39d0071d998459b101f2bd1e13f09193bc5a4 +file antiref96/train/shard-00071.parquet 118844390 7f6ce754a2085381ac0ad44e4b4fc7bc09cd970f03d87372fdbfc2a54f96756d +file antiref96/train/shard-00072.parquet 118681135 020b7c7d9c68edb9a84a3e2e9cbfc641b27566a15b37e6e7070d2170eae4c6d3 +file antiref96/train/shard-00073.parquet 118423047 5a4c544e298c7d5f77c2e71745717480cb69f99be3a045660c82cfc7ca6adcb7 +file antiref96/train/shard-00074.parquet 118908184 3e2746c021789d16616e0fc7b17adca06dd637e1888f93b9a1a2e91dc8d6a871 +file antiref96/train/shard-00075.parquet 118432761 db6085f7eb670209a0278793281edddff19336faeb71beef7a6609e222547518 +file antiref96/train/shard-00076.parquet 118542894 c1c69799fa7f5bd1786a494fe554dd07ec6aaee22f032b6b982ff38706c58f32 +file antiref96/train/shard-00077.parquet 118662920 d75787a9657d6ef20f32bd938d5b2046e3853f4fdf04ceea0228fb93eead9cf5 +file antiref96/train/shard-00078.parquet 118160280 ce67f6ee4a6f8d4cbf6bf6f92d0b26dff6331c448c92fed0a180bc15442aa7c8 +file antiref96/train/shard-00079.parquet 118533164 3302b7d9ac0f7936f5e1481b902b1e41b45b96171e53499b5bfce14800421695 +file antiref96/train/shard-00080.parquet 118255025 63b0f02e2805c11584f8bb93595960347b476551b46c78ce6f1f09d0d259e21b +file antiref96/train/shard-00081.parquet 118695686 038c7ecb8071d113b4a4c05b99bffec7897eccf1811ccd64a25a9c8310c08006 +file antiref96/train/shard-00082.parquet 118955854 6db594fded36440b8b86ada72119bbd7667d4b3629eef84653585f679e92fe6b +file antiref96/train/shard-00083.parquet 118185592 64e73050287c91322e37386662981ba04f2ad811e64c65a4068cafde6b611ef1 +file antiref96/train/shard-00084.parquet 118805361 66ea46ef5f48d17e28765b4f629bd6e9a139d03b68778cf5988de0df78fb8a22 +file antiref96/train/shard-00085.parquet 118989314 26a4c0ca5ac069e9e560dce164fffe507475f5eee9d6571c4700aaaa984ad601 +file antiref96/train/shard-00086.parquet 118218923 82c9d232828af01019c4cdb9343bc4564e31737af0b8218920bc4ef8b9777bdc +file antiref96/train/shard-00087.parquet 118717346 1a7ec674ed79fab12350e0320b050cb909747d603182245d68c5ba775e9986a5 +file antiref96/train/shard-00088.parquet 118997885 6ab145c9f532396c25cfeff5758a0bf10d65ac9ea6f81143020cd3a7cb686c3c +file antiref96/train/shard-00089.parquet 118651273 b289d688d6ef09e3913776e7f7090bda9e6297b870911932743c22f13ea00089 +file antiref96/train/shard-00090.parquet 118464579 1548b93d56a8171ef31bff10ebd52c980900d847366af6fcbbf07e54972d2fbf +file antiref96/train/shard-00091.parquet 118415626 f93ed10c3bd13bcc3e259969f77539a5fa530bf3ee479b6494069fdcd31a5138 +file antiref96/train/shard-00092.parquet 118588947 cdbc75a04768997644b648760fb3d8c26c71c22d25374ea77076e204ffe2305c +file antiref96/train/shard-00093.parquet 118300722 b9ce4e389d28adcd854ee88490045d39786a75bdbafa5d62ad037d7ef1978b5f +file antiref96/train/shard-00094.parquet 118249070 14326341eb84f936de702f7d0f9bd44fb9c12d73c9fa755007d5e7a46f3abe7e +file antiref96/train/shard-00095.parquet 118512792 19bfd26ca9f2573f4fe88ce58d957677de0f27977aa3baa7fd83997d8ec16d35 +file antiref96/train/shard-00096.parquet 118851707 d3644bf66cd35229af5a71707bf6147bc67f0edd4500ab76efe3b1431abc88ef +file antiref96/train/shard-00097.parquet 118614839 8ebfe724759a87bf719747c138b98892397eb360bfa9f958f8e1c6e0429e667b +file antiref96/train/shard-00098.parquet 118569406 6e9cdb4798f7d26f5d07d623aca9b947501c5a87f00eb7fe9d287d7bee70b603 +file antiref96/train/shard-00099.parquet 118017076 27a9eba7badf0ff7592c0f5db35db3364270e03a91bb29aff2f37adae7ddb8be +file antiref96/train/shard-00100.parquet 118819682 81e1fc51d1f23404e2efe427c5b0a41f81ba39aa6246becc913e6276bcad9844 +file antiref96/train/shard-00101.parquet 118730584 3268e16be04a0e900c531b97225b8859df27922d0bfa85ab2522a624ade8c40f +file antiref96/train/shard-00102.parquet 118013968 507346bbb8f1a14c747a8286bdbaa0f79709872aad57247a609db1d2707507f3 +file antiref96/train/shard-00103.parquet 118589615 282ce82abbee1aed982ccceffeb6958149a858c57d83bc57a9e6ee973311ea10 +file antiref96/train/shard-00104.parquet 119124217 725df5ea30e476a435eb9aea619bcd561bd261c383bdcd1a1f5e36093a49c471 +file antiref96/train/shard-00105.parquet 118558529 be0c39fbae7beed47d3df1516f2b46f07c68969d045fb0b1e85ed8a0b148d4f2 +file antiref96/train/shard-00106.parquet 118647636 5da9e3aa18221e485ac2e9337e3bee167860bff4f80aade7cadd5bc51ea32194 +file antiref96/train/shard-00107.parquet 119073417 78390b753e5bc222e7645c1816f9d68678db669050f4a4c0d4e163db9a000e4a +file antiref96/train/shard-00108.parquet 118464827 da77a21d55ba97b5d522d4c344712c8d6bcdb2d1888c5e1ca2e7767ff2f9aefa +file antiref96/train/shard-00109.parquet 118509346 ff054cbce880828756c2b73c6074c22d9fa3c9b725c6caa637f178c8fef1a141 +file antiref96/train/shard-00110.parquet 118170377 b5f1342715d035f79354f3d2ec86c386e1753f3ef6a3906b9e1d8b6c1914dcf3 +file antiref96/train/shard-00111.parquet 118635223 56948570eca090178c1b018519801d8862c3c62f3ca75ac72aec78204aca0f93 +file antiref96/train/shard-00112.parquet 118350679 0f35f408f4767dd0e79b90125687360cd351f8dec9707162d35ba7482777d6b0 +file antiref96/train/shard-00113.parquet 117946284 8a6899c48d59b4a9c88308cc5cadbdccc4e0b9fe2d078b72f68c60e9e08f8c6d +file antiref96/train/shard-00114.parquet 118152824 6d1e3af16436fda8ec57c7efabf011dc0e3bcd8dd02bca8b9ce450da488eb2ae +file antiref96/train/shard-00115.parquet 118411099 1b59eb5a75602af103e9e3f34dd4d63b6796fdb5a090c9356181270673c3ac98 +file antiref96/train/shard-00116.parquet 118635747 1acfb000ac1586814758a5c88989572e8afb04ea239c82c719ba07d676812de1 +file antiref96/train/shard-00117.parquet 118570880 381230374c48c47511e13ee6250a7a55bc5f06cf86b3403a6ebc385a5eb569cc +file antiref96/train/shard-00118.parquet 118419309 a6d781046756088694a12de31fccfa48f5210c4545d97a9a31af4025fbbcd9ae +file antiref96/train/shard-00119.parquet 118452679 c0b9f62946daafe224d58a92a28ff4162ee04ae2355af33e6969570cf0b67e79 +file antiref96/train/shard-00120.parquet 118242567 7c6e1e8d0544c0133f39bb2ea3d2f244a9df7407a6fde93722cd74033df533a2 +file antiref96/train/shard-00121.parquet 118320321 d461f329f9f792542a2e2ae6313e5b4c3773801b83dbcacb4b29b59e94d49dd7 +file antiref96/train/shard-00122.parquet 118866470 9ff81048074fd520ffe21c965fbad4117a191dadc9e38c348fb6b7ce05ad8131 +file antiref96/train/shard-00123.parquet 118520330 d6deb1da5fbbb6a874470e33f187c219ed29c5531aea35c3ca745c7ac67a220f +file antiref96/train/shard-00124.parquet 118685071 c8dad2ed3d03658ac799bdfcde4ae4c2b0c2b958a0a42b5d2eeab66ae818e064 +file antiref96/train/shard-00125.parquet 118149335 6a8a23f2aa5e23d790016fe15afbe4cd3493ba7cbf2e705bb96c8e8af512201c +file antiref96/train/shard-00126.parquet 119198774 e2e013da1e7ed36d6f02342ccda4c8caea768b528fb4996eb32ff32003c8e145 +file antiref96/train/shard-00127.parquet 118429632 508e065bcc025b9c447b63f6e0ddb4c6c5b581c0cb708444e98dbfb0f1989e42 +file antiref96/train/shard-00128.parquet 118495534 6e9582eb7b659ccaf74367c4155c2a12545ecedc8c7ecd9e872d37cc79d004ad +file antiref96/train/shard-00129.parquet 118586106 e203638c0b29b19c8b15560658248ed0896be874b786993a87fead3a35cd121b +file antiref96/train/shard-00130.parquet 118462531 f49abad06a82adaa18c1984e8635361639c6dea7a1b5c66aeaa84b9a1e1f198f +file antiref96/train/shard-00131.parquet 118304885 192b262bac080d3ee1bd1c0f305a2aee0e7c56e0895663442d82a6f7f72c8309 +file antiref96/train/shard-00132.parquet 118325098 7014c9c85bac7cc043143b9cf3e900505d1fbea714f441e4c24ce2029ea98463 +file antiref96/train/shard-00133.parquet 118598858 9584e7b7e08e3b1fb42e557a229251c9a20a959718362e5e54dfe6a9b4e1c251 +file antiref96/train/shard-00134.parquet 118658566 55b570ce7f32eddb5555e6fc7f4434dbbded349f40c43dbcaff8cb673a8424c3 +file antiref96/train/shard-00135.parquet 118901211 296287f70bedb66bd2dd19ebc592f724760dae1745adc9df972d488509cb3bf3 +file antiref96/train/shard-00136.parquet 118376936 d6924c2684b72055a08dca293d6f67a04291ed49b07e9a9f433a059f9ecc8fc6 +file antiref96/train/shard-00137.parquet 118560895 08b2aff4743c4ac479771f23c2e53cf765efe60df56eb3ca75d02699fbe65331 +file antiref96/train/shard-00138.parquet 118340835 08a38cf8ff0a2c7b94df505cc2729d23d5f3a7c1dd375241c6fb44c9d5aedee9 +file antiref96/train/shard-00139.parquet 118296844 c184334e66348605559c274a8e8e393eca2b91383e5a49b23fe74779c095f090 +file antiref96/train/shard-00140.parquet 118462270 e443a4a39a369886b07c1c45facc09d8aa9de0b75850177f4221efa1f94b7c5a +file antiref96/train/shard-00141.parquet 119266606 05ec816a5459ce82109bf7e1cfdc933e6365804be2f27e06eadfb48b2b68ca1e +file antiref96/train/shard-00142.parquet 118594785 9bd54233e69f94c066157ae349934e49212da2e367cf0ace4f934478d747d6ae +file antiref96/train/shard-00143.parquet 118544144 794ba16b06c4b0953dafeef3b4eb5400e488280dd290df45067721472cc66962 +file antiref96/train/shard-00144.parquet 118461727 fd3c7f2e933fb9b116acf73a219a72b5a76e13220573ea2ecb3a132e107b40b2 +file antiref96/train/shard-00145.parquet 118422437 8493ef06071dbd607b4c32db3959945589a95f06a124cb7d217c3928ab4a34ae +file antiref96/train/shard-00146.parquet 118350877 d8eab727903116139907291347d67208600db5d4e7736a192ab87ba05b5978b6 +file antiref96/train/shard-00147.parquet 118924005 ef808dbcd19ca296321d63b7e5a2f223f0b3e714726e07573ee66efac37778e5 +file antiref96/train/shard-00148.parquet 118583713 dcaf0f51de184c27384aedd8fbda93989a342f838fb60405605cc3ee73ab7802 +file antiref96/train/shard-00149.parquet 118968487 5e820b531ec173d1383bbf26aba9b09fa2affc30f2be2a068011601d5ca253ce +file antiref96/train/shard-00150.parquet 118206408 d6b4d067150899b8d7bd8e141f86caf19005aaef6327dfa7864c9795a7b69814 +file antiref96/train/shard-00151.parquet 118188739 e27989d935db29141a40df6c717c83f2dbd7ab1cc2a2f881a35fb01824bff8e7 +file antiref96/train/shard-00152.parquet 118205673 b18b2b92d4e36268e7563ce44a5dbbf27baa53460f135596d28b6d1a5ce909bf +file antiref96/train/shard-00153.parquet 118376815 1b625c3219a3f049f7fea22994688b6fd98264fac6fd601dc93d4a2ca1470548 +file antiref96/train/shard-00154.parquet 118550235 bf93852ac611dc14ec0d5975684f0a76c6a96cbde7086358fd04fbe5627bbb7b +file antiref96/train/shard-00155.parquet 118459515 e01d9502e7d0200d0d82a10d2099846b1578206c96bf5f705d8edcdd761b0247 +file antiref96/train/shard-00156.parquet 118122334 fdec4647378d5cdc51c143f178b5521bf28c7beccc68be5f245137e9ea433e62 +file antiref96/train/shard-00157.parquet 118669109 616531657002029cdba8b9758a7b034a93559551cf7233622cff0b9b0daa2743 +file antiref96/train/shard-00158.parquet 118347011 7e9988743e399fff18c3daa10db8d795e36590f9aa5433036444d63ec163af8b +file antiref96/train/shard-00159.parquet 118721241 4fdb49ca0213e94a66e6756c67b106806af1ae2bb335eeffec4d22346a6cad69 +file antiref96/train/shard-00160.parquet 118357124 b2cb78639b29b9556c8c767132ba67162d047053ad6f17c8c8681a2f04b60c9a +file antiref96/train/shard-00161.parquet 118590492 aafc977780825453c397fc685bcfdc0dd897890d76c448f5de3a88f2dbd5e4e1 +file antiref96/train/shard-00162.parquet 118626970 bab89fdb9ad6a3c54e7c244fefda252ff735ed0e3d1b84177091a393317b2689 +file antiref96/train/shard-00163.parquet 118460731 cfd6c196464d6d6373797a40304350442cab41ae2379d667507ec391a3b05b1f +file antiref96/train/shard-00164.parquet 118586384 ae99dbbc9fb35b69db8014213029798a661311e5b46c0c2eb3fb94b9dc380de9 +file antiref96/train/shard-00165.parquet 118519058 d02033077a7ca824acc1ea01fe0cc5a49570e78984bac97fc19ada4295920d52 +file antiref96/train/shard-00166.parquet 118977585 c5d24ace383270429217eb4badcde226e124e0e8f71c0a3bf581acb2293c9fbf +file antiref96/train/shard-00167.parquet 118286464 c5b7ccb27db709e006b3f330d994ce2eb6f38ead9f5a7c77f4ca42ae6b1554ff +file antiref96/train/shard-00168.parquet 118694823 2614edebd5ea55b1416ab7269f26fc247bed040fcc1479776e3bc34fb1d672e7 +file antiref96/train/shard-00169.parquet 118226659 746264872f9c7105b86ac577f8ef731de03633641ed0eb03dc65de8b45ed9768 +file antiref96/train/shard-00170.parquet 118952843 00988f50ec380ab54127b11b9a6f2e2b424e8b697b7d093e11485306ad03ed83 +file antiref96/train/shard-00171.parquet 118879286 f0a5ab43a851eccb536f8a4aa8e2f3415651086feecebe276fa29e140a94b571 +file antiref96/train/shard-00172.parquet 118543285 4c7968e3bd12b98aa1377acd3e16f02e864afa366ded7ef7618daf0d5eb82c3d +file antiref96/train/shard-00173.parquet 118806645 c64532c18f9bef2350d10da569e70946ee031583fc948ffee6a576e68c66a139 +file antiref96/train/shard-00174.parquet 119168567 09a248ec82e69e96b6c56f29648d2ea214d4439174a766a33b9d1acfbf29aea4 +file antiref96/train/shard-00175.parquet 118262851 37654b81515c002620c3b04a0f5895d83a1bbcf55a58c50475170ef30dfc31bb +file antiref96/train/shard-00176.parquet 118321162 184adb62eac7156404156c7b9bee427029b54690deeec02d3fe3aa4a269f0142 +file antiref96/train/shard-00177.parquet 119084244 47f9a8abb8449f183a7b6c77e2a41197232b9263bdfda253c0f19fa11f38a1e5 +file antiref96/train/shard-00178.parquet 118489181 482de7df10feb1d7ecbc82dd5e894e17fd3643bb885b12c26d02ee893a5dfa35 +file antiref96/train/shard-00179.parquet 119013922 10e1c9ab382546d9493e1da81b35341a75c320a03095e8e0eb17efc266f286f4 +file antiref96/train/shard-00180.parquet 118629745 1fcbb362df65eb5c9ae5d52fdebcb85ba2c5cf92355ef03372e878f5fe87a816 +file antiref96/train/shard-00181.parquet 118150670 40c53e0b777427021a309696dd4f09da5ca20527ca30b0b02a380dcbc00d7bb8 +file antiref96/train/shard-00182.parquet 118988664 16f414e071a3094c602ad57526c9b4cafbb66cd05218e3b33f626d48202b1edf +file antiref96/train/shard-00183.parquet 118615729 d69c381efe51c797ac5bce64532603994da3db53254989caef0d14c846b45e01 +file antiref96/train/shard-00184.parquet 118071834 59a5e1cd63617864430437c676525d0b8579243c2145fb422fdb59253044488e +file antiref96/train/shard-00185.parquet 118881692 468fef9844ddd2269148e71acbf3c75069649fbd17e24c9b9239c6f5d8396ba8 +file antiref96/train/shard-00186.parquet 118770751 cc9a58e2627247530c29b3a5723242fe97581bc136288fdceb4876b3d76856c4 +file antiref96/train/shard-00187.parquet 118612769 5b8fa816d4cebb0113914a90da60f32b860bc90ba39e5e80d8566c163ee668a8 +file antiref96/train/shard-00188.parquet 118471102 6adeb374333274c0070b44d31fb444db4775746b8e4bd150967c5e9929352b91 +file antiref96/train/shard-00189.parquet 118382816 ba63415f0b1241f6aa7f226b330b2651fa6c2d063be8929e7d72b5f2bed10066 +file antiref96/train/shard-00190.parquet 118539583 c546fad4b22707e82834bd923f7e1ab6ff994ca499a2f9f6b9df5e4181d80727 +file antiref96/train/shard-00191.parquet 118770284 d99e0e611b9f107a2793195031fc1cc69eb7dbb8a37f3b3fe6662a4a38ce0c02 +file antiref96/train/shard-00192.parquet 118387030 86d488ca23e2ed79600c3e136accf9b5ed4d5063b40744c24fda0fe64335f1b4 +file antiref96/train/shard-00193.parquet 118749430 1db61280829cb6a4be1408e9d09b865abf748f2296244937477bcf24cefe6df8 +file antiref96/train/shard-00194.parquet 118497527 cf29102c9b34bcb8869e87c8454c2a1315a49f51542e9cabc7698cfbf97ab61e +file antiref96/train/shard-00195.parquet 118748940 21537db6ef730a5bca6abb27132a0acb8dc01f96e4cac3afcec590779d7825c6 +file antiref96/train/shard-00196.parquet 118461750 e72b51738fef2ce8bb8370628f83e8f3f13fe84c586686a086bfcf79053e9a7b +file antiref96/train/shard-00197.parquet 118806887 4d748e44bceb8e2612fe6daf56eb43e698e28a70d6bba0fba8756a12a9f3a216 +file antiref96/train/shard-00198.parquet 118461881 9cd9da9baa3ac848412b226a77a9cc2813449c6e1f66395514c04fc3243e18ce +file antiref96/train/shard-00199.parquet 52611825 81aa77eff26e4a96dbf8ffccf247f7f933f8bfc12116665049432746a94eccc6 +file eval_cdrh3.parquet 133772 9381cd7bb0160d205c1943f2efbf0d6dd3054a0bc4be3898dd47629afe602898 +file eval_gte2.parquet 259750 862f73fe8c1faadcf98c50c5b6c83742dac6e6fc36c1b1331ee4896e6ee28c99 +file eval_singletons.parquet 261751 f28027d3b65fc11e1ac790674d5be1885775973c072da9dfd4e4e91899648552 +file test_cdrh3.parquet 133101 29c801bd78e74e9e4941c962c51fffe12081118847392c547e0de755aab6bbdc +file test_gte2.parquet 257271 9456f11b468f016492fc58f03817de55a8b385879a73ded08ee50a92073aa9d5 +file test_singletons.parquet 261764 cc6381e9e0fedb8c447925c9d369ef57e219237ab61f60055166b27ed8688758 +commit: 08de241378cb9ce06370926d52cf811a78dbdc6c