TRL Training on Hugging Face Jobs

Train language models using TRL on managed Hugging Face infrastructure. Models train on cloud GPUs and results save to Hugging Face Hub automatically.

Key Rules

1. Submit via hf_jobs() MCP tool — use hf_jobs("uv", {...}) with inline Python code in the script parameter. Local file paths do not work (jobs run in isolated Docker containers). When a user asks to train a model, create the script and submit immediately.
2. Include Trackio in every training script for real-time monitoring. Use example scripts in scripts/ as templates.
3. Enable Hub push — the training environment is ephemeral; unpushed results are lost. Set push_to_hub=True, hub_model_id, and secrets={"HF_TOKEN": "$HF_TOKEN"}.
4. Set adequate timeout — default 30 min is too short for most training. Add 20-30% buffer to estimated time.
5. Validate unknown datasets before GPU training to prevent format failures. See references/dataset_validation.md.
6. Report after submission — provide job ID, monitoring URL, estimated time. Wait for user to request status checks.

Training Methods