Release v0.5.3

Highlights

• Support using AutoModelForSequenceClassification.from_pretrained load reward model @xiaoxigua999 a9c482a
  • The default value_head name of the reward model has been changed to score
• Support RLOO with per-token KL penalty by @zhuzilin in #515

What's Changed

• Fix bug on prm trainer w.r.t no packing samples and ring attn by @zhuzilin in #551
• Fix issue #549 ZeroDivisionError during DPO/SFT/RM/KTO/KD training eval step by @MarxistZ in #561 @xiaoxigua999

New Contributors

• @MarxistZ made their first contribution in #561

Full Changelog: v0.5.2...v0.5.3