Skip to content
Kitbay

trl-training

Apache-2.0Updated on Nov 11, 202611,334GitHub stars

A compact Hugging Face skill for post-training language models with the TRL library, through its Python API or the trl command line.

What the skill teaches

A compact Hugging Face skill for post-training language models with the TRL library, through its Python API or the trl command line. It maps each training method (supervised fine-tuning, DPO, GRPO online reinforcement learning, distillation, KTO and reward models) to its trainer class, config and expected dataset format, explains the fields that matter for SFT and the setup for GRPO reward functions, notes that configs extend the Transformers training arguments, and points to the experimental trainers. It is guidance only and links to the TRL documentation for more.

Capabilities

  • Pick the TRL trainer and dataset format for a method
  • Set up supervised fine-tuning runs
  • Configure GRPO reward functions
  • Run training from the trl CLI