model-trainer

Fine-tuning LLMs with SFT and DPO on Hugging Face Jobs is handled through TRL-based training workflows. The skill supports end-to-end training by requiring a YAML frontmatter, a training script template, and hub push configuration to execute cloud GPU jobs.

The best way to monitor TRL training experiments is using Trackio-based monitoring. The skill integrates dataset validation and Trackio monitoring to ensure reliable results and debuggable experiments during your Hugging Face Jobs runs.

Yes, you can use PEFT with TRL for GRPO and reward modeling workflows. The skill dependencies include peft and trl to enable end-to-end training workflows on Hugging Face Jobs.

To convert a fine-tuned model to GGUF for local deployment, you follow the skill's GGUF conversion guidance. This provides instructions for local deployment and CPU/GPU inference after your TRL training job completes.

Yes, you need Hub authentication and push-to-hub workflow configuration to save your training results. The skill requires this setup to enable seamless Hub persistence and preserve your fine-tuned model outputs.

Running TRL training jobs without dataset validation risks unreliable experiment results. The skill includes dataset validation as a core feature to ensure reliable, debuggable experiments and prevent errors during SFT, DPO, and GRPO workflows.