# Training code bundle for the trait-10 (curiosity) 716 arm

| field | value |
| --- | --- |
| `experiment` | Code-only bundle for a credential-free RunPod pod that trains the trait-10 difficult-advice arm: the trainer, its src/ import closure and the train config. The training DATA is not here -- the mixture is a public dataset the trainer resolves through its normal data_repo path, sha-pinned in the config. |
| `date_generated` | 2026-08-20 |
| `constitution` | none -- this repo holds code. The constitution the training data connects to is declared on the mixture repo the config pins. |
| `source_repo` | https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ d1fa94d14499b20f35215269b5a86ee43fb5eded |
| `models` | trains a LoRA adapter over Qwen/Qwen3.6-27B |
| `generation_config` | {"files": ["pyproject.toml", "scripts/train/train_lora.py", "src/__init__.py", "src/utils.py", "src/huggingface.py", "src/model_profile.py", "src/train/__init__.py", "src/train/train_lora.py", "src/train/masking.py", "src/train/mask_gate.py", "src/train/dynamic_batching.py", "configs/train/2026-08-25_lora_qwen36_table2_9284_difficult_advice_716_verbose_dynbatch.yaml", "configs/train/2026-08-25_lora_qwen36_table2_9284_difficult_advice_verbose_token_matched_dynbatch.yaml"], "train_config": "configs/train/2026-08-25_lora_qwen36_table2_9284_difficult_advice_716_verbose_dynbatch.yaml"} |
| `schema` | code.tar.gz -- a tarball to extract at the repository root. |
| `provenance` | uv run python scratch/trait10_curiosity/train_pod.py bundle |