huggingface / trl

The skill's own description names the training methods TRL supports today — supervised fine-tuning, DPO, GRPO, KTO, and reward-model training — plus the trl command-line tool for running SFT and DPO jobs without writing code. It points an agent at the same trainer classes the README's own quick-start examples use, which should keep an agent's code suggestions in step with a library that ships new trainers and even a new major version, TRL v1, fairly often.

This is not a general skill library the way most packages here are — trl is a machine-learning research codebase with one skill grafted onto it, not a curated set. Reaching for it only makes sense if you are already running or about to run TRL training jobs; there is nothing in it for anyone who has not installed TRL itself and does not intend to.

READ THE FULL ANALYSIS＋

It has not been verified to run. Our dependency check on this skill came back unverified, which means we have not confirmed that its setup actually installs and executes cleanly end to end — treat it as a pointer to read, not a tested integration.

Who it is for. Anyone fine-tuning or aligning an open model on Hugging Face's own stack. If your training pipeline runs on a different library — Axolotl, a standalone Unsloth script, a custom PyTorch loop — this skill's API references will not transfer, because it is scoped entirely to TRL's own trainer classes.