Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
MoritzLaurerย 
posted an update 2 days ago
Post
1426
The TRL v0.13 release is ๐Ÿ”ฅ! My highlight are the new process reward trainer to train models similar to o1 and tool call support:

๐Ÿง  Process reward trainer: Enables training of Process-supervised Reward Models (PRMs), which reward the quality of intermediate steps, promoting structured reasoning. Perfect for tasks like stepwise reasoning.

๐Ÿ”€ Model merging: A new callback leverages mergekit to merge models during training, improving performance by blending reference and policy models - optionally pushing merged models to the Hugging Face Hub.

๐Ÿ› ๏ธ Tool call support: TRL preprocessing now supports tool integration, laying the groundwork for agent fine-tuning with examples like dynamic temperature fetching in prompts.

โš–๏ธ Mixture of judges: The new AllTrueJudge combines decisions from multiple binary judges for more nuanced evaluation.

Read the release notes and other resources here ๐Ÿ‘‡
Release: https://github.com/huggingface/trl/releases/tag/v0.13.0
Mergekit: https://github.com/arcee-ai/mergekit
Mixture of judges paper: The Perfect Blend: Redefining RLHF with Mixture of Judges (2409.20370)
In this post