Sandy-sys/aditi-ps18-state

ADITI PS-18 — programme state

Working state of the OSINT fine-tuning programme on Google's Gemma 4 12B QAT checkpoint: adapters, corpora, gate reports, training code and the session log. No secrets here.

Start with STATE_2026-08-21.md — current state, open blockers and the resume plan. Then SESSION_LOG.md and RESTORE.md for the programme up to 17 Aug 2026.

Adapters

LoRA r=64 on q,k,v,o,gate,up,down × 48 layers unless noted. Optimizer states are not uploaded (2.1 GB each, not needed to reuse an adapter).

adapter	schema	ner F1	sent	claim	halluc	faith	gen %	note
v8full_bf16_r64	0.950	0.764	0.855	0.894	0.364	0.834	54.5	only run holding schema and faithfulness
v9_r64	0.960	0.563	0.230	0.862	0.302	0.159	53.5	halluc bucket over-corrected to 29%
v10_r64	0.195	0.740	0.830	0.893	0.546	0.000	63.5	faithfulness collapses here
v11_r64	0.855	0.747	0.820	0.886	0.369	0.245	61.0	schema_lock bucket raised to 9.5%
v12_firm / v12_gentle	0.505 / 0.470	0.769 / 0.773	0.890 / 0.870	0.903 / 0.905	0.523 / 0.479	0.027 / 0.000	48.5 / 40.5	repair corpus, warm-start v10
v13_a	0.820	0.755	0.885	0.897	0.476	0.000	18.5	lr 5e-5, MCQ collapse
v13_b	0.905	0.756	0.875	0.881	0.500	0.053	64.0	best composite of 22 runs; lr 9e-5
qat_r64_WRONGGRID_amax7.0	—	—	—	—	—	—	—	⚠️ see below — kept for the record only

200 items/task, bare decoding, no constrained grammar. schema is the 11-key OSINT record; faith is grounded-answer schema validity. Every run from v12 onward is warm-started from v10_r64, which is how the faithfulness regression propagated into v13_b.

Two things worth knowing before you reuse any of this

1. W4A16 export erases the structured-generation capability. v13_b goes 0.905 → 0.505 on schema_lock after int4 export; v8full_bf16_r64 goes 0.950 → 0.000. The export is not at fault — 328 packed modules matching Google's w4a16-ct exactly, zero module-set difference, merging verified lossless. It is genuine int4 damage, worst on long nested objects.

2. qat_r64_WRONGGRID_amax7.0 was trained against the wrong grid — do not use it. It fake-quantised with scale = amax/7.0. compressed_tensors uses amax/(bit_range/2), and int4's range is [-8,7], so bit_range = 15 and the divisor is 7.5. Measured on real layers, that mistake put the training grid 8.6% away from the deployment grid — larger than the 6.7% quantisation error the adapter was being trained to absorb. With 7.5 the two grids match bit-for-bit (rel-err 0.000000). The corrected qat_train.py is in code/.

Layout

STATE_2026-08-21.md   current state, blockers, resume plan
adapters/             LoRA weights + per-run metrics.jsonl
corpus/               v10, v11, v13, repair, sentiment corpora (gzipped) + bucket reports
reports/              every gate report, one JSON per run
logs/                 training and scoring logs, ship pipelines
code/                 trainers, hybrid merge, W4A16 export, the contract + harness

Datasets are referenced by repo id + commit SHA, never redistributed — 207 sources, 169 pinned. Reproduce with python refetch.py --manifest sources.json.