# LoRA SFT adapter — recipe `sft` on mixture `da-gpt-fluff-removed-7` (qwen36, seed 0)

| field | value |
| --- | --- |
| `experiment` | LoRA SFT adapter — recipe `sft` on mixture `da-gpt-fluff-removed-7` (qwen36, seed 0) |
| `date_generated` | 20260915 |
| `constitution` | inherited from the training data (dougalldeepmind/2026-09-15-da-gpt-fluff-removed-7-mix); not declared at launch |
| `source_repo` | https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 9a7d93b3e443b533e81098ba31d6b87149f5fce9 |
| `models` | base: Qwen/Qwen3.6-27B@6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 |
| `generation_config` | {"recipe": "sft", "seed": 0, "thinking": true, "epochs": 1.0, "lr": 0.0001, "batch_size": 1, "grad_accum": 16, "max_seq_len": 8192, "lora": {"r": 64, "alpha": 128, "dropout": 0.05}, "dynamic_batching": {"token_budget": 8000, "loss_agg": "seq-mean-token-mean"}} |
| `schema` | PEFT LoRA adapter (safetensors) + tokenizer + train_config.yaml (the resolved config that ran, every launch argument and pin written back: `uv run train --config train_config.yaml` re-runs it) + training_meta.json {organism, thinking, recipe, mix_subject, train_config, base_model, base_model_revision, model_profile, dataset{repo,file,revision}, git_sha, timestamp} |
| `provenance` | scripts/train/train_lora.py --config configs/train/sft.yaml model=qwen36 data_repo=dougalldeepmind/2026-09-15-da-gpt-fluff-removed-7-mix data_file=mixture.jsonl data_revision=5444733e6cd26222516cae093c33cd332822c7b2 base_model_revision=6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 seed=0 hf_repo=2026-09-15-qwen36-da-gpt-fluff-removed-0 wandb=true |
| `dataset` | hf.co/datasets/dougalldeepmind/2026-09-15-da-gpt-fluff-removed-7-mix@5444733e6cd26222516cae093c33cd332822c7b2 (mixture.jsonl) |