| license: other | |
| tags: | |
| - rl-forgetting | |
| - coding-sft | |
| # Fresh coding SFT runs, November 2, 2026 | |
| Thirteen freshly executed arms from the October24 request:Qwen2.5-3B and Qwen3-1.7B-Base, each ordered/shuffled/uniform/hard/CE;Llama-3.2-3B-Instruct uniform/hard/CE. The two Llama ordered/shuffled baselines were newly run October27 and are linked in request-provenance.json. September training results are historical and are not counted as fresh execution. Partial transfer is not completion. Check delivery_manifest.json for verified training units and evaluations/mbppplus_n160_t0.6/delivery_manifest.json for execution-scored evaluations. | |
| lr1e-5,one epoch,seed42,effective batch256,micro1,4H100,4096tokens,10checkpoint saves per arm. Replay lambda1.0. Models/checkpoints are under runs/checkpoints/step_N/model and runs/final in corresponding model repositories. Exact inputs,source,retained recovery state and metrics are in this artifact repository. | |
| Qwen2.5-3B derivatives retain Qwen Research License;Qwen3-1.7B-Base derivatives retain Apache2.0;Llama3.2 derivatives retain Llama3.2 Community License and Acceptable Use Policy. See model LICENSE/NOTICE/USE_POLICY. No extra rights to upstream models are granted. | |