OSS Contributions
My open source work is mostly in the ๐ค Hugging Face ecosystem โ TRL and PEFT โ along with Halo and curated datasets published on the Hub.
โ merged ยท ๐ in review
Featured: GOLD Trainer in TRL
Extending TRL's GOLDTrainer (Cross-tokenizer On-and-Off-Policy Distillation):
Other contributions to TRL
- โ
Fix
chunked_nllpatch hiding VLM kwargs fromgenerateโ PR #6156 - โ
GRPO: treat explicit
max_tool_calling_iterations=0as zero, not unlimited โ PR #6423 - ๐ GRPO + RLOO: feed vLLM unexpanded image placeholders (vLLM VLM fix 1/2) โ PR #6718
- ๐ GRPO + RLOO: un-expand tool-returned images (vLLM VLM fix 2/2) โ PR #6719
- ๐ GRPO + RLOO + Distillation: fix multi-process tool loops in vLLM server mode โ PR #6673
- โ
Respect
use_cpuincreate_model_from_pathdevice-map default โ PR #6295
PEFT
- โ Documentation improvements โ PR #3406
โช Halo
Contributions to Halo, White Circle's open-source framework for training large language and multimodal models:
- ๐ Generalized JSD(ฮฒ), reverse KL, and teacher top-k distillation losses โ PR #106
Datasets on the ๐ค Hugging Face Hub
- MMK12-QWEN35-9B-VERIFIED โ a verified teacher-trace dataset for vision-language model distillation.
- LLaVA-CoT-filtered โ a cleaned 87k-example LLaVA-CoT variant with answer-leakage removal.
