Epic: round 7 — Unsloth training and dynamic requant, served via Ollama, scored on a fresh three-slot benchmark

[Xore]

Why this round exists

#1947 is about to finish measuring every model as shipped on this card. Phase 2 settled on 2026-09-06 (49 measured / 3 unmeasured of 52), phase 3's clean f16 ladder (#2245) completed at 10:26Z, and the cold re-measurement of the whole roster now runs on this round's pin (see the status block below). When that lands, the as-shipped question is answered.

Two things that matrix already shows, and neither can be fixed by pulling one more tag:

1. The top is saturated. Sixteen models sit within one point of each other at 63–64 / 69 on Tier B, from a 12 B dense up to a 35 B MoE; the incumbent qwen3:14b is at 62. The self-quantised gemma-4-26B-A4B ladder scored 62 / 62 / 62 at Q3 / Q4 / Q5 — bits do not move the score, the model does, and on a 14-case rubric a real training gain would be invisible.
2. The 27–35 B class spills. The best rows (Ornith-1.0-35B Q4_K_M at 22 GB, the Qwen3.6/3.8-27B class at 17–19 GB) run partly on the CPU at ctx 32768: 6.7 min/run against 2.1 for a resident model doing identical work. The published quants stop where our card cannot use them.

This epic is the other axis: train our own models for our own three jobs with Unsloth, quantise the bigger bases ourselves with imatrix + per-tensor bit selection so they fit, serve everything through the pinned Ollama, and score it all on a fresh three-slot benchmark with headroom. Ollama does not train and cannot produce imatrix quants; Unsloth does not serve. The plan is the seam between them.

Plan of record: docs/benchmarks/plans/2026-09-06-round7-unsloth-train-requant-ollama.md (PR #3089). Read it before touching any child issue; it carries the measured status, the hardware envelope, the pipeline, the data rules, the run order and — §15 — how agents work it.

Decisions taken 2026-09-06 (operator)

question	decision
Unsloth capabilities in scope	all four: SFT/QLoRA of the slot models; RL (DPO then GRPO) with the benchmark scorer as reward; continued pretraining on domain text; dynamic requant (imatrix + per-tensor bits)
"bigger models"	requant the 27–35 B class to GPU-resident and the 100 B+ RAM-offload class (123 B / 218 B rows); GLM-4.6 357 B stays a measured hardware rejection
training data	a separate, decontaminated corpus; real captured data from ES allowed as input but never leaves the host; captured slices use a local teacher; synthetic slices may use open-weight frontier teachers via OpenRouter (licences permit distillation)
compute	local card only, queued behind the cold re-run; training above 14 B only at the VRAM edge (offload) or as a gated follow-up
harness for round 7	main's 17-case / 83 rubric + pooled-claims scoring + all three slots, one new pin for the whole round; the a99e765 / 69 numbers are historical context only

Rules that carry over from #1947 (not relaxed)

• One scoring vintage per table. Round 7 gets one pin; state it in every report.
• Cold slot or the numbers are contaminated. STOP_WORKERS=1, N=2 → 3 → 5 escalation, UNRESOLVED marker, VRAM sampler, uptime per run, weights kept above the free-space floor.
• 0 ≠ empty. UNMEASURABLE / UNMEASURED are markers, never scores.
• Transcripts are the deliverable. Synthetic runs commit them under docs/benchmarks/runs/; captured-data runs keep them outside the repo.
• Never train on the test set. The 17 corpus programs — any toolchain, any optimisation level, their decompiled text, their rubric terms — are off limits for training, calibration and RL prompts. A decontamination report is a deliverable, not a footnote.
• The GPU is one card. No training container starts while a sweep is alive; every chain waits on a positive condition, never on "the GPU looks idle".
• Nothing lives only in /mnt-1. Scripts, configs, manifests and run cards are committed; adapters and merged weights are mirrored.

Working conventions — grit, rtk, gh (plan §15)

• grit is mandatory for parallel work. One agent name per subagent (issue-<N>-coder). grit gc → grit symbols --file '<area>/*' → grit plan -a … -i "…" → grit claim -a … <file>::<symbol> … (creates .grit/worktrees/<agent>/; edit only the claimed symbols, only there; --with-deps, --queue as needed) → grit heartbeat --ttl 900 on long legs → grit done -a … (serialized; never two at once). Claim before dispatching a coder. New files: claim the neighbours you touch and grit init after the merge. Absolute paths inside grit worktrees. grit init again after any large merge to main.
• rtk on every shell command. The hook rewrites commands to their rtk form; do not bypass it. rtk proxy <cmd> only when raw output is genuinely needed; long outputs to a file, read selectively. rtk gain --history in every end-of-issue report.
• gh conventions unchanged. Check gh issue view and gh pr list --search before claiming; in-progress + assignee + comment; milestone comments; push → PR → CI green → gh pr merge <n> --squash --delete-branch; Closes #N; new issue for out-of-scope findings; audit issue state at batch end; no AI tooling mentioned anywhere; no credentials or production domains in issues.
• At most two agents at once, and only one of them on the GPU, through the positive-condition chain.

Status 2026-09-06 afternoon — the #1947 cold re-run was folded into this round

Operator decision: instead of 2–4 GPU-days re-measuring ~97 tags on the saturated 14-case pin, the whole roster is measured once, cold, on the round-7 pin. Done today (STATE-2026-09-06-round7-fold.md on the homeserver; scripts analysis/ghidra/benchmarks/corpus/round7_*.sh on PR #3089):

step	result
old-pin cold re-run	stopped after 13 models / 52 files → 1947cold/ kept as valid OLD-pin cold cells, ABORTED-2026-09-06.txt, partial quarantined
phase 3 (#2245)	completed 10:26Z; ornith-35b-selfquant:Q4_K_M = 0 on all four runs → UNMEASURABLE on that pin, re-measured here
pin	second clone APIARY-round7 detached at 32dbdeb1; the a99e765 clone untouched, transcripts mirrored
corpus + cache	17-case corpus rebuilt in the provenance container, manifest byte-identical; tierb-cache-round7/ 17 entries, 0 errors, Ghidra 11.3.2
smoke	qwen2.5:7b-instruct-q4_K_M 70/83 A, 69/83 B, 17 cases, 0 empty — max is 83, not 79
running since 14:08Z	round7_coldrun.sh → sweep_extra.sh: 96 tags, cold, N=2 → 3 → 5, results round7/, log round7.log, VRAM sampler beside it; 2–4 days

Order

running    round7_coldrun.sh — the ghidra-slot cold baseline of this round on pin 32dbdeb1 (leave it alone; 2–4 days)
no GPU     #3080 toolchain ▸ #3082 corpus v1 ▸ #3081 merge/export proof (CPU) ▸ #3086 calibration sets ▸ #3087 slots + chain drivers
GPU frees  #3080 smoke ▸ #3081 scoring ▸ #3086 imatrix + ladders ▸ #3083 CPT ▸ #3084 SFT ▸ #3085 DPO/GRPO
then       #3087 scores every artefact on the same pin, all three slots, cold ▸ write-up ▸ promotion records
gated      #3088 training above 14 B

Children

• round7-1: Unsloth training toolchain on the homeserver — pinned container, Ada-only GPU, merge → GGUF → imatrix → Ollama export path, smoke test #3080 — round7-1: Unsloth training toolchain on the homeserver — pinned container, Ada-only GPU, merge → GGUF → imatrix → Ollama export path, smoke test
• round7-2: pipeline proof T0 — REx86 adapter → 16-bit merge → GGUF → Ollama, scored against its base and #160's PEFT numbers #3081 — round7-2: pipeline proof T0 — REx86 adapter → 16-bit merge → GGUF → Ollama, scored against its base and Evaluate the REx86 adapter through a supported exact-base local runtime #160's PEFT numbers
• round7-3: training corpus v1 — decontaminated, teacher-labelled slices for the three slots, injection pairs, CPT text; manifest + decontamination report #3082 — round7-3: training corpus v1 — decontaminated, teacher-labelled slices, injection pairs, CPT text; manifest + decontamination report
• round7-4: continued pretraining with Unsloth on domain text — decompiler output, sessions, RE corpora — before SFT #3083 — round7-4: continued pretraining with Unsloth on domain text before SFT
• round7-5: SFT QLoRA of the slot students (Qwen3-14B, gpt-oss-20b, Gemma-4-12B) on corpus v1 — per-slot and multi-slot adapters, exported at equal bits #3084 — round7-5: SFT QLoRA of the slot students (Qwen3-14B, gpt-oss-20b, Gemma-4-12B) — per-slot and multi-slot adapters, exported at equal bits
• round7-6: preference and RL round — DPO then GRPO with the benchmark's own judges as reward (schema, injection gate, polarity, pooled claims) #3085 — round7-6: DPO then GRPO with the benchmark's own judges as reward
• round7-7: dynamic requant ladders — imatrix + per-tensor bit selection for the 27–35B class and the 100B+ offload class, against #2245's plain ladder #3086 — round7-7: dynamic requant ladders — imatrix + per-tensor bits for the 27–35 B class and the 100 B+ offload class, against Research: quantize bigger bases to fit the Supermicro — is large-model-low-bit a real gain over the current slots? #2245's plain ladder
• round7-8: benchmark round 7 driver — 17-case /83 pin, pooled claims, all three slots, cold protocol, chained behind the cold baseline, write-up and promotion records #3087 — round7-8: benchmark round 7 driver — 17-case /83 pin, pooled claims, all three slots, cold protocol, chained behind the cold re-run, write-up and promotion records
• round7-9 (gated): training above 14B on this card — 27B dense QLoRA at the VRAM edge with Unsloth offload, or a recorded rejection #3088 — round7-9 (gated): training above 14 B on this card, or a recorded rejection

Dependencies: 3080 → {3081, 3083, 3084, 3085, 3086}; 3082 → {3083, 3084, 3085}; 3081 green before 3084 scores; 3084 → 3085; 3087 consumes every row; 3088 gated on 3084's result and the operator.

Still needed from the operator

1. OpenRouter key on the homeserver in a 0600 env file (S3 labels only; never in the repo).
2. 2 × 16 GiB DDR4 registered ECC RDIMM into DIMME1 + DIMMF1 — not a blocker; makes the 100 B+ rows and activation-offload training faster.
3. Private Hugging Face repos for adapters — default off.
4. round7-9 (gated): training above 14B on this card — 27B dense QLoRA at the VRAM edge with Unsloth offload, or a recorded rejection #3088 opens only on your call after round7-5: SFT QLoRA of the slot students (Qwen3-14B, gpt-oss-20b, Gemma-4-12B) on corpus v1 — per-slot and multi-slot adapters, exported at equal bits #3084's first result.

Related

#1947 (the matrix this round extends), #2245 (the plain K-quant ladder = this round's control), #1804 (domain-specialised candidates), #2279 / #2985 (phase 2.5 and the still-missing slots_sweep.sh — #3087 absorbs the slots sweep), #356 / #1523 (the REx86 adapter path #3081 finally exercises), #2969 (Ollama 0.33.3 requalification — separate, but the newest archs may need it), #3023 (why the cold protocol is in the driver).

[Xore]

Prepared and running — the #1947 cold re-run is folded into this round (2026-09-06)

Stopped: the a99e765 cold re-run after 13 models / 52 files (1947cold/, kept, ABORTED-2026-09-06.txt); phase 3 had already completed (#2245: gemma-4-26B-A4B ladder flat at 62–64; every Ornith self-quant UNMEASURABLE — Ollama 500s, harness bug #3090).

Prepared, all committed on PR #3089 as analysis/ghidra/benchmarks/corpus/round7_*.sh:

pin	APIARY-round7 detached at 32dbdeb1; a99e765 clone untouched, its transcripts mirrored
corpus	17-case corpus rebuilt in debian:trixie-slim via ci_verify.sh, manifest byte-identical, 280 semantic executions 0 failed
Tier B cache	17 entries, 0 errors, Ghidra 11.3.2
smoke	qwen2.5:7b-instruct-q4_K_M 70/83 A, 69/83 B, 17 cases, 0 empty — max is 83, not 79 (title/body/plan corrected)
driver	sweep_extra.sh + GHIDRA_CACHE / OPERATOR overrides; round7_coldrun.sh (pin + cache guards); round7_launch.sh (refuses without a Tier B smoke)

Running since 14:08Z: 96 tags, cold, N=2 → 3 → 5, results round7/, log round7.log, sampler beside it. First cell qwen3.5:4b Tier A 71/83. Expect 2–4 days; the training legs queue behind it on the positive condition.

Orchestrator handoff: docs/benchmarks/plans/2026-09-06-round7-hermes-init.md (same PR) — prep mode, GPU gate, kickoff order, grit + ponytail + rtk + gh conventions, dispatch pattern, hard rules. Live record on the host: /mnt-1/benchmarks/STATE-2026-09-06-round7-fold.md.

Still owed by the operator: OpenRouter key on the host (S3 labels only), the two DIMMs (not a blocker), the private-HF decision (default off), and the word START.

[Xore]

Correction to the comment above: PR #3089 (the plan) was merged before the drivers were pushed, so the drivers, the §0 status block, the /83 correction and the orchestrator handoff are on PR #3091 (feat/round7-drivers). Round-7 leg verified on the real run: qwen3.5:4b Tier A 71/83 ×2, Tier B run1 75/83 over 17 cases.

[Xore]

**Tracker: #3106 (glyphsoftware/sentinel-r3) is NOT covered by this round yet — adding it here so it doesn't get lost.

Verified 2026-09-08:

• Weights exist on the Supermicro at /var/training/models/sentinel-r3 (raw safetensors, 15 shards) — but there is no GGUF conversion and no Ollama entry (api/tags on ghidra-ollama-1: 177 models, zero sentinel). Not pulled to Ollama, not served, not smoke-tested.
• No sentinel-r3 row in any One benchmark, every model, every aspect: rebuild the model matrix on current hardware with error bars #1947 matrix, no runs under docs/benchmarks/runs/, no PR references benchmark: add glyphsoftware/sentinel-r3 to the model matrix #3106.
• Round-7 driver round7-8: benchmark round 7 driver — 17-case /83 pin, pooled claims, all three slots, cold protocol, chained behind the cold baseline, write-up and promotion records #3087 slot lists and the cold baseline do not include it.

Per #3106 acceptance this needs: GGUF quant (Ollama serving per the corrected serving note — vLLM retracted), Tier A zero-token smoke, cold scoring on the 17-case /83 pin via the #3087 driver, results recorded in the #1947 matrix. Raw safetensors alone do not satisfy any of it.

[Xore]

Round7 dependency disposition — #3083 / #3084

SFT-only is not waiting for completed CPT. Plan §7 and #3084 define T2 against the untouched base, T3 against T2, and T2′ from T1 where CPT finished. The GPU sequencing diagram is not a universal artifact dependency. This correction permits documentation/acceptance preparation for T2/T3 now; it does not authorize execution or change the queue.

What exists: toolchain/corpus/driver deliveries (#3092, #3093, #3116, #3117), T0 export metadata consistent with #3136, surviving CPT artifacts and stored result metadata. Fresh host-side metadata also shows S4 licence annotations on 5,981 split records and 37 S6 manifest rows with hash fields. These are not complete lineage or split-integrity certificates.

Actual gates, by path:

• T2/T3: acceptance for the selected SFT slices, shared toolchain acceptance, T0 green before scoring, and the positive-condition resource handoff before any GPU launch. feat(benchmarks): round 7 Unsloth training toolchain #3092 explicitly deferred smoke; feat(benchmarks): corpus v1 — synthetic slices, OpenRouter labelling, decontamination scan #3093 deferred full real-corpus acceptance. CLOSED issues do not establish these deferred deliverables.
• S4-consuming paths (revdeck per-slot and the multi-slot union): the surviving real slice report says S4 38 hits, pass=false. Preserve that failure until the corpus owner identifies the exact scanned input and any existing accepted replacement report. No false-positive ruling or waiver made. Other slots still need their own accepted data; S4 is not automatically their dependency.
• T2′ additionally: authoritative CPT input/base/artifact lineage and independent durability evidence. The S6 report records zero hits, but its binding to the actual historical training input remains unestablished; it does not certify S1–S4.
• T0: the export manifest is not behavioral acceptance. Root-level UNMEASURED markers do not prove no later T0 results exist. The acceptance owner must identify the existing behavioral acceptance record and required postings, or record that artifact unavailable—not launch a rerun on this assumption.
• round7-4: continued pretraining with Unsloth on domain text — decompiler output, sessions, RE corpora — before SFT #3083 historical documentation: measured training tokens/s and peak VRAM remain distinct from reported inference throughput. Missing measurements must stay UNMEASURED if unrecoverable. The recorded configuration deviation needs an explicit experimental disposition, not silent tuning. Neither gap creates a universal CPT dependency for T2/T3.

Next bounded step: prioritize a sanitized SFT-only, per-slot acceptance packet; request exact corpus and T0 acceptance references from their owners. Keep T2′ conditional. Queue owner separately owns a current roster/status + process/lock/worker handoff; filenames or an idle GPU do not establish it.

These findings are present-day metadata reconstruction (host timestamp 2026-09-17T01:49:51Z), not invented historical measurements or certification of missing lineage. No safety, provenance, licensing or decontamination requirements waived. No samples/transcripts inspected, no model loading/training/inference/scoring or service changes. Any subsequent training needs a separate scope review; nothing launched. Issues remain open/unchanged by this disposition.

[Xore]

Round 7 tracking gap closed — evidence handoff, 2026-09-17

The existing September 17 readiness/S4 reviews are now reconciled in the per-slot handoff. This supplements, rather than replaces, the 01:55:53Z dependency disposition.

Newly explicit blockers:

• Ghidra and sessions: known empty selected inputs. Operator verified each train/dev file at 0 bytes / 0 lines. S1/S2 source maps are empty; S3 has 63 rows, 63 unlabelled, kept=0. This is stronger than the earlier filename-only readiness finding, not evidence of absence everywhere.
• Revdeck and union: slice report records 4,045 train / 575 dev each; operator verified revdeck train 4,045 lines / 1,599,265 bytes. No retained S1–S3 contribution is established by the source metadata. Nonempty outputs are not accepted multi-component data.
• S4: surviving September 14 scan metadata remains 4,642 scanned / 38 hit events / pass=false. Source report has zero train exclusions and 22 dev exclusions (597→575). No exact input binding; 38 events are not necessarily 38 samples. Neither residual final-training hits nor an accepted replacement follows from these counts.
• Preserve the September 9 zero-hit train-only and calibration claim. It predates the surviving failure and is not final SFT train/dev acceptance. Dates, populations and acceptance scopes must remain distinct.

Existing evidence is not erased: T0 export metadata/result-file leads exist. CPT reconciliation preserves historical 74/83 A twice, 75/83 B twice; 32–33 tokens/s is inference, not measured training throughput. #3083 stays open for historical lineage, independent durability, configuration disposition and missing measurements, which remain UNMEASURED if unavailable. Only T2′ adds CPT-specific acceptance; SFT-only T2/T3 do not universally require completed CPT.

Idle is not released: the supplied operator observation at 2026-09-17 22:51:46+02:00 found no matching workload processes and GPU 0%, 2 MiB. That does not certify roster completion, process/lock/worker ownership or a resource handoff. Model-mirror timer was disabled and service failed/not running; neither was changed. PR #2641 was checked and remains open/draft. #3085/#3088 remain gated; closed delivery issues #3080/#3081/#3082/#3086/#3087 are not blanket acceptance certificates.

Next bounded action: corpus, toolchain/T0 and queue owners attach existing dated acceptance references or explicit unavailable dispositions to the linked packet. Required: per-slot identities/approvals/split and scan binding; existing smoke/T0 acceptance and postings; current positive-condition handoff. CPT owner supplies the additional lineage/durability record only for T2′. No new scan, corpus repair, benchmark, training, copying, service change, closure/reopening or waiver. Any execution needs separate scope authorization and applicable acceptance/resource gates. No automatic successor work.

[Xore]

T0 references reconciled — 2026-09-17

Requirement-by-requirement disposition on #3081. This updates only the toolchain/T0 reference gap in the earlier handoff.

• Required postings exist: September 9 full table on #1523 and #356 T0 closing summary. The September 7 posting-pending statement was stale. Sixteen cold runs are reported, 17 cases, N=2 consistent: adapter Q4/Q8 A 68/68, B 62/65; matched base A 64/70, B 61/64. Existing metadata independently supports result-file presence; no scores recomputed.
• Still UNVERIFIED in the bounded sources: complete shared training/export smoke and exercised template-diff/release evidence; result-bound runtime/scorer/export provenance and committed-run references; the four Evaluate the REx86 adapter through a supported exact-base local runtime #160 behavioral comparisons, injection verdicts, and adapter/base per-run wall-clock; dated acceptance against those clauses. Exports and totals alone do not settle those requirements. Evaluate the REx86 adapter through a supported exact-base local runtime #160's native-runtime acceptance and older Wire the approved REx86 adapter into a real Rev·Deck production path (GGUF or native PEFT) #356 harness comparisons are different vintages, not automatic T0 acceptance.
• Next: toolchain/T0 owners attach the existing sanitized artifacts and clause-level disposition, or state precisely which artifact is unavailable. No presumed rerun, approval, waiver, closure, reopening or successor agent. No material new toolchain-specific finding warrants duplicating this on round7-1: Unsloth training toolchain on the homeserver — pinned container, Ada-only GPU, merge → GGUF → imatrix → Ollama export path, smoke test #3080.

Corpus findings remain as documented, without re-audit. SFT-only does not universally depend on CPT. Resource handoff stays a separate live gate; idle GPU is neither readiness nor authorization. PR #2641 remains open/draft with its merge hold. No source, service, resource or model-backup changes; no execution authorized.

[Xore]

T0 host-record search complete — 2026-09-18

First direct host-side record search performed (read-only, metadata-only). Recovered: full export-manifest bindings (base revision, adapter/ZIP digests, six GGUF SHA-256, merge/quant method), 16/16 per-run provenance (recorded model digests, unique transcript run IDs + digests, uniform seed-144 qualification, N=2 identical stored totals and per-case scores), all 272 stored per-case wall_seconds (per-run inference wall-clock sums recomputed, labeled), stored #160-named comparison and injection verdict metadata, and resolution of the September 9 retry logs (pre-registration "model tag not installed" attempts, superseded by the successful runs).

Searched and unrecoverable: the 16 transcript artifacts themselves (four independent search methods, zero hits), the #3080 training-smoke packet and container exit record, and any historical per-run harness/image identity field. Full detail and exact search scope: #3081 (comment)