All terms
Frameworks & Tools
TRL
A Hugging Face library of trainers for fine-tuning and aligning language models.
Definition
TRL is a Hugging Face library that provides trainers and utilities for the post-training and alignment of language models. It implements supervised fine-tuning, reward modeling, and methods that teach a model from human preferences (such as PPO, DPO, and GRPO) on top of Hugging Face Transformers and PEFT (tools for cheap fine-tuning). It is a standard toolkit for assembling RLHF-style pipelines, where models learn from human feedback.