Skip to content

oplm-antibody-sequences 2.0.0 - #16

Merged
briney merged 15 commits into
mainfrom
oplm-antibody-cdrh3-design
Nov 14, 2026
Merged

briney merged 15 commits into
mainfrom
oplm-antibody-cdrh3-design

Conversation

@briney

@briney briney commented Nov 14, 2026

Copy link
Copy Markdown
Member

Cook and publish oplm-antibody-sequences v2.0.0: a heavy-chain CDR3-identity holdout at 85%, plus four screened train sets (AntiRef90, 92, 94 and 96 clusters with AntiRef99 members), for briney/oplm#78's comparison of antibody train sets.

Why. 1.0.0 held out whole AntiRef90 clusters only, so 27% of eval_gte2's heavy chains had an exact copy of their CDR3 in train. 2.0.0 adds 2,500 + 2,500 held-out heavy chains whose CDR3s have no 85%-identical CDR3 left in any train set. 1.0.0's four eval and test files carry over byte-identical, and the test checks their sha256.

What changes from 1.0.0

  • New: eval_cdrh3.parquet, test_cdrh3.parquet, antiref90/, antiref92/ and antiref94/train.
  • antiref96/train is now screened. When the screen took a cluster's representative, its first surviving member leads the row.
  • antiref99/train is dropped.
  • Train results are not comparable across the two versions.

Build (record: results/oplm-antibody-sequences/2.0.0.txt; Hub candidate commit 08de2413)

  • 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: 152.4M (1.97)
    • AntiRef92: 161.8M (1.85)
    • AntiRef94: 175.1M (1.71)
    • AntiRef96: 199.4M (1.50)
  • All four folders hold the same 299,748,717 sequences.
  • About 7.5 h on bcr with 32 CPUs. The screen took about 5.7 h of that.
  • The test reruns the screen on sampled members with its own matching code (DuckDB levenshtein plus a q-gram-filtered semi-global search) and passed.

PI decisions along the way, recorded in the spec and skill

  • 85% CDR3 identity at any length, with 2,500 + 2,500 held out (2026-11-12).
  • Four clustering levels in one dataset: 1.0.0's D3 exception, extended (2026-11-12).
  • The tripwire raised from 5% to 6% after the first build removed 5.134%. The held-out sets were kept as drawn (2026-11-13).
  • A new version of this dataset rather than a new dataset, without antiref99/train (2026-11-13). This also departs from D3.

The branch also carries the design (docs/superpowers/specs/2026-11-12-oplm-antibody-cdrh3-design.md) and plan (docs/plans/oplm-antibody-cdrh3.md). After review: gh pr merge --merge, then scripts/promote oplm-antibody-sequences 2.0.0.

🤖 Generated with Claude Code

briney and others added 15 commits November 11, 2026 21:17
…% from AntiRef, with AntiRef90/92/94/96 train sets for OPLM's anneal-mix sweep

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…f, and the unanchored fallback's bound is floor(0.25 x query length)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ooked on bcr, then briney/oplm#78's four-branch sweep

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…DR3s at 80% identity by a pigeonhole index, with a self-check

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nct CDR3s and screens every AntiRef99 train sequence against them

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…creened AntiRef99 sequences, promoting a surviving member when the screen took the representative

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t's 80% separation, and the screen on sampled members

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… 2,500 + 2,500 held out, a 5% tripwire, after measuring the screen at 80%

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…distance, holds out 2,500 + 2,500 heavy chains, and trips at 5% removed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ops the build

The first bcr build removed 5.134% (> 5%) and threw away a ~6 h screen. The hits
now land in .work/screened.parquet before the check, so a rerun with a raised
screen_max only re-checks the share.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ild removed 5.134%

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…est holds each train folder to 90% of that build's rows

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…creened train sets, replacing the separate oplm-antibody-cdrh3, without antiref99/train (PI)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@briney
briney merged commit 672504c into main Nov 14, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant