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

[Xore]

Part of #3079 (round 7). Plan §4 and §7, experiments T2 (per-slot) and T3 (multi-slot). GPU work; queues behind the cold re-run.

Students, and why these

The 20 GB envelope decides. Unsloth's published minimums: QLoRA 4-bit 14 B ≈ 8.5 GB, 27 B ≈ 22 GB, 32 B ≈ 26 GB; 16-bit LoRA 7 B ≈ 19 GB. MoE bases in 4-bit QLoRA are not recommended by Unsloth (bitsandbytes gap) and their 16-bit LoRA needs 60 GB+ — so gemma-4-26B-A4B, Ornith-1.0-35B (qwen3_5_moe) and Qwen3.6-35B-A3B are not trainable on this card; they are round7-7's requant targets instead. The one MoE that does fit is gpt-oss-20b through its native MXFP4 path (~14 GB).

student	why	fit
Qwen/Qwen3-14B	the incumbent in all three slots (qwen3:14b, 62 / 60) — a fine-tune of the exact production weights is the cleanest possible A/B	≈ 8.5 GB QLoRA
gpt-oss-20b	the only 20 B-class model trainable here; also the #2279 harmony-shape family that scored 0 as shipped — worth knowing whether training fixes it	≈ 14 GB MXFP4
Gemma-4-12B	the OBLITERATED variant is joint top of the matrix at 64 / 66 from a 12 B — the best small base; pick the derestricted variant as the start point (refusals on malware evidence are a failure mode here) and record which	≈ 8 GB QLoRA

Use Unsloth's dynamic 4-bit variants (unsloth/<model>-unsloth-bnb-4bit) where published — they exist precisely to lose less in QLoRA.

Runs

• T2: one adapter per slot per student — ghidra (Tier A + Tier B evidence shapes, S2 + S3), sessions (S1), revdeck (S3 + S4) — train_on_responses_only, r = 16–32, α = 2r, lr 2e-4, 1–3 epochs, packing, seq 8 k (32 k for the ghidra adapter with unsloth_tiled_mlp=True), use_gradient_checkpointing="unsloth", fixed seed; early stop on corpus v1's dev slice by schema-valid rate and dev rubric, not by loss
• T3: one multi-slot adapter per student on the union — production runs one model per slot today, but a single model serving all three is a real deployment option and costs nothing extra to measure
• T2′: the same as T2 starting from round7-4's CPT checkpoint, for the students where CPT finished
• every run exported through export_to_ollama.sh at Q8_0, Q6_K, Q4_K_M and an imatrix IQ4_XS (round7-7's calibration set) so the fine-tune is compared with its base at equal bits, never confounded with quantisation

Deliverables

• train.py + configs under analysis/ghidra/training/, run cards under docs/benchmarks/training/runs/ (loss, dev metrics, VRAM peak, wall-clock, tokens/s)
• adapters + merged 16-bit under /mnt-1/training/runs/t2-*/, mirrored; optional private HF adapter repo (operator decision)
• every artefact registered in round7-8's row list with its base-at-same-quant control
• a Qwen3 note in the run card: thinking is off in the production contract (thinking: false); train non-thinking data with enable_thinking=False in the template and verify the exported Modelfile carries the same template — the Bump esbuild from 0.25.12 to 0.28.1 in /dashboard/frontend in the frontend-compatible group #1 cause of degraded exports per Unsloth's docs

Depends on round7-1, round7-2 green, round7-3. Feeds round7-6 and round7-8.

[Xore]

Working conventions for this issue: grit (claim the symbols you touch with grit claim -a issue-<N>-coder … where N is this issue number, edit only in .grit/worktrees/issue-<N>-coder/, grit done to merge, grit init after the merge if you added files), rtk on every shell command (rtk proxy only for raw output; rtk gain --history in the closing report), and the gh claim/comment/PR/merge conventions. Full text: epic #3079 → "Working conventions", and the plan's §15.

[Xore]

Claimed for round 7 (orchestrator, 2026-09-06). Pre-flight done: plan of record merged via #3089; grit initialised at repo root; no GPU work starts until the cold re-run completes.

[Xore]

Round 7 evidence handoff — 2026-09-17

Reconciles the existing readiness review (17:21–17:22Z), S4 follow-up (18:11–18:12Z), and operator-supplied aggregate checks from September 17. This is evidence transfer, not a new scan or execution approval. Supplements the September 17 dependency disposition.

Per-slot disposition

Path	Established input evidence	Remaining gate
T2 ghidra, S2+S3	Both train/dev files operator-verified 0 bytes / 0 lines; report agrees. S2 source maps empty. S3 has 63 rows, all 63 unlabelled, kept=0.	Known empty selected inputs, not merely missing filenames. Accepted S2/S3 data packet unavailable in reviewed evidence. S4 and completed CPT are not automatic dependencies.
T2 sessions, S1	Both train/dev files operator-verified 0 bytes / 0 lines; S1 source maps empty.	Known empty selected inputs; own accepted S1 packet still needed.
T2 revdeck, S3+S4	Report: 4,045 train / 575 dev. Operator confirms train 4,045 lines / 1,599,265 bytes. S3 contributes zero kept rows.	Nonempty is not accepted: unresolved S4 scan-to-input binding, selected-source/composition and split acceptance.
T3 union, S1–S4	Report: 4,045 train / 575 dev; listed composition does not establish populated components. Source metadata has no retained S1–S3 contribution.	Component/composition acceptance missing, plus S4 gate. Equal counts do not establish byte identity or a complete multi-slot corpus.
T2′, conditional CPT starting point	Existing CPT artifacts/results are documented separately.	Selected-slot/shared gates plus accepted historical CPT lineage and independent durability. Not a universal T2/T3 gate.

S4: preserve the failure and the historical claim

• Surviving scan, file mtime 2026-09-14 08:10:25Z: 4,642 scanned, 38 hit events, pass=false, 499 protected documents. No input hash/reference fields establish its binding to final train/dev. 38 events need not mean 38 distinct samples.
• Slice-report mtime 2026-09-14 08:10:26Z: S4 train 4,045 rows, kept=4,045, dropped=0; dev 597 rows, dropped=22, kept=575. All 22 exclusions are dev. The arithmetic 4,642−22=4,620 is consistency evidence, not lineage or proof of residual training hits.
• The September 9, 17:08:02Z historical claim reports train-only 4,185→4,045, then 0 hits on 4,045, and calibration acceptance. Credit that dated claim; it predates the September 14 failure. The existing calibration report (6,881 scanned / 0 hits, September 9 mtime) and S6 report (37 / 0, September 11 mtime) do not certify final SFT train/dev acceptance. The historical train-only claim does not cover the later 575-row dev population.

Neither final-input contamination nor an accepted replacement is established by these captures. Mtimes are not execution dates or causal proof. Prior searches were bounded/truncated, with a permission error; no universal absence claim is warranted.

Concrete administrative deliverables — existing references only

1. Corpus/acceptance owner (round7-3: training corpus v1 — decontaminated, teacher-labelled slices for the three slots, injection pairs, CPT text; manifest + decontamination report #3082): dated per-slot packet with exact final train/dev manifest hashes, aggregate label coverage/composition, licensing/teacher/locality approvals, group-level split certificate and protected-set/scanner identities. Link the failed scan's input identity and any existing post-filter report to those final inputs; supply an authorized dated acceptance/supersession disposition preserving the original failure. If unavailable, say unavailable; no waiver or repair is implied.
2. Toolchain/T0 owners: existing end-to-end smoke acceptance and T0 behavioral acceptance plus required postings. #3092, #3093, #3116, #3117 establish deliveries, not all deferred acceptance. T0 export metadata and 16 result-file leads exist; neither filenames nor root UNMEASURED markers settle behavioral acceptance.
3. Queue owner: dated roster/result-or-UNMEASURED reconciliation, process/lock/worker ownership and release acknowledgment, exclusive next owner/window and resource/storage guard acceptance. The operator check at 2026-09-17 22:51:46+02:00 found no matching workload processes and GPU 0%, 2 MiB; that is a point-in-time observation, not resource handoff or absence of all artifacts.
4. CPT owner, T2′ only: existing lineage/durability references and explicit historical configuration-deviation disposition. September 17 reconciliation preserves stored 74/83 Tier A twice; 75/83 Tier B twice. The approximately 32–33 tokens/s is inference, not training throughput. Missing historical training token count, throughput and peak VRAM remain UNMEASURED if unrecoverable.

Next bounded action: owners attach existing sanitized references or explicit unavailable dispositions to this packet. Keep #3083 open and #3085/#3088 gated. No closure/reopening, waiver, new benchmark, corpus remediation, training, model copying or service changes authorized by this handoff. Any execution requires separate scope authorization after applicable acceptance and current queue gates.

[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)