LASR-Callum/2026-08-31-qwen36-lora-table2-9284-difficult-advice-principle-scoped-702-rank-64-seed-69

Seed replicate (69) of the principle-scoped 702 difficult-advice arm — the project's difficult-advice BASELINE (docs/BASELINES.md). Table2 9,284 filtered + 702 difficult-advice rows (7.03%) whose two refine stages were never shown the whole constitution. Trained so the baseline carries seed-to-seed variance: with one seed it could not be ranked against any sibling.

field	value
experiment	Seed replicate (69) of the principle-scoped 702 difficult-advice arm — the project's difficult-advice BASELINE (docs/BASELINES.md). Table2 9,284 filtered + 702 difficult-advice rows (7.03%) whose two refine stages were never shown the whole constitution. Trained so the baseline carries seed-to-seed variance: with one seed it could not be ranked against any sibling.
date_generated	2026-08-31
constitution	claude_distilled_12_principles_mid (9 principles). NOTE this arm's generator saw only ONE principle per stage — the constitution's preamble (the priority / conflict-resolution section) reached it through no stage at all.
source_repo	https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT — training config configs/train/2026-08-31_lora_qwen36_t2_9284_da_chunk_only_702_dynbatch_s69_2xh200.yaml (its pre-rename name, as run)
models	Base Qwen/Qwen3.6-27B; corpus written by anthropic/claude-haiku-4.5 + anthropic/claude-sonnet-5 via OpenRouter.
generation_config	seed 69; LoRA r=64 alpha=128, lr 1e-4 cosine, warmup 0.05, global batch 16, 1 epoch, dynamic batching, max_seq_len 8192, 2xH200 DDP, 625 steps. Identical to seed 0 but for seed/output_dir/hub_model_id.
schema	PEFT LoRA adapter + training_meta.json (thinking stamp + the sha-pinned dataset).
provenance	scripts/gpu/runpod_train.py up --bundle LASR-Callum/2026-08-31-da-chunk-only-702-seeds-bundle then scratch/da716_seeds/push_adapters.py. Data pinned at LASR-Callum/2026-08-31-da-chunk-only-702-seeds-bundle@818256371774 :: t2_9284_da_chunk_only_702.jsonl.