0z5a c36846c8a4 fix: condition online DPO on rollout prompts ...

- Concatenate rollout prompts with selected chosen and rejected responses
- Mask prompt tokens from DPO loss while preserving explicit attention visibility
- Cover response selection, padding alignment, and prompt-conditioned inputs

2026-09-02 12:52:50 +08:00

..

__init__.py

chore : 升级 torch 2.11.0+cu128，移除自定义 Muon，修复 gloo device_id

2026-06-27 16:10:37 +08:00

metric_util.py

refactor: deduplicate low-risk code paths

2026-08-19 16:17:40 +08:00

rollout.py

refactor: split infer core into subpackages by concern

2026-08-08 23:43:05 +08:00

schedule.py

refactor: remove dead code and deduplicate scheduler setup

2026-08-19 14:58:37 +08:00

strategy.py

fix: condition online DPO on rollout prompts

2026-09-02 12:52:50 +08:00

train_callback.py

refactor: rename optimizer step callback hooks to before and after

2026-09-01 14:22:54 +08:00

train_context.py

feat: load HuggingFace checkpoints via key/config conversion

2026-08-20 11:34:59 +08:00

trainer.py

refactor: rename optimizer step callback hooks to before and after

2026-09-01 14:22:54 +08:00