| # Artifact index |
|
|
| This repository is a curated debugging handoff. It is not a claim that any |
| included adapter improves the base model. |
|
|
| ## Checkpoints |
|
|
| | Directory | Base | Configuration | Why included | |
| |---|---|---|---| |
| | `checkpoints/4b-base-dpo-cleanprompt/step-28` | `Qwen/Qwen3.5-4B` | DPO, LR 1e-6, beta 0.1, LoRA r16/a32 | Early checkpoint with the best held-out partial-reward signal | |
| | `checkpoints/4b-base-dpo-cleanprompt/step-125` | `Qwen/Qwen3.5-4B` | Same run | Final checkpoint, showing late degradation | |
| | `checkpoints/9b-dpo-length-clean/step-8` | `prism-drift/qwen35-9b-m0-v4` | DPO on 245 length-clean pairs | Early dev improvement | |
| | `checkpoints/9b-dpo-length-clean/step-12` | `prism-drift/qwen35-9b-m0-v4` | Same run | Checkpoint associated with held-out reversal | |
| | `checkpoints/9b-grpo-probes/default-step-16` | `prism-drift/qwen35-9b-m0-v4` | LR 1e-6, beta 0.05, LoRA r16/a32 | Default probe | |
| | `checkpoints/9b-grpo-probes/aggressive-step-16` | `prism-drift/qwen35-9b-m0-v4` | LR 5e-6, beta 0.01, LoRA r16/a32 | LR/beta probe | |
| | `checkpoints/9b-grpo-probes/r32a64-step-16` | `prism-drift/qwen35-9b-m0-v4` | LR 1e-6, beta 0.05, LoRA r32/a64 | Higher-capacity probe | |
|
|
| Each checkpoint contains only: |
|
|
| - `adapter_model.safetensors` |
| - `adapter_config.json` |
| - `trainer_state.json` |
| - the generated checkpoint `README.md` |
|
|
| No optimizer, scheduler, RNG, duplicated reference adapter, tokenizer copy, or |
| merged full-model cache is included. |
|
|
| ## Supporting evidence |
|
|
| - `evaluations/`: exact completion-level artifacts and summaries used in the |
| report. |
| - `data/`: final 4B Base DPO pairs, the 4B/9B M0 length-clean subsets, and the |
| final Base sweep/cohort manifests. |
| - `metadata/runs/`: immutable run metadata and available checkpoint-time reward |
| logs. |
| - `logs/`: selected formal-run and hyperparameter-probe logs. |
| - `configs/`: training configurations used during the investigation. |
| - `code/`: the relevant training, filtering, merge, and evaluation utilities |
| as they existed when this package was assembled. |
| - `reports/rl_training_debug_report.md`: the full narrative report. |
|
|
| `SHA256SUMS` covers every uploaded file other than `SHA256SUMS` itself. |
|
|
