huggingface-llm-trainer

The huggingface-llm-trainer skill guides training and fine-tuning of language and vision models on Hugging Face Jobs infrastructure using TRL (Transformer Reinforcement Learning) or Unsloth, without requiring a local GPU. It solves the problem of running managed cloud training end to end: models train on cloud GPUs and results are automatically pushed to the Hugging Face Hub. It covers SFT, DPO, GRPO, and reward modeling, plus GGUF conversion for local deployment with Ollama, LM Studio, or llama.cpp.

Its capabilities include submitting jobs via the hf_jobs MCP tool with Python scripts passed directly as strings, using UV scripts with PEP 723 inline dependencies, and including Trackio for real-time monitoring in every training script. It provides guidance and example scripts for SFT, DPO, GRPO, and Unsloth training, dataset preparation and validation, hardware selection with a detailed flavor table (T4, L4, A10G, A100, and TPU options) mapped to model sizes and budgets, cost estimation, and memory optimization techniques such as LoRA/PEFT, gradient accumulation, gradient checkpointing, and mixed precision. Reference files cover training methods, hub saving, Trackio, GGUF conversion, macOS local training, Unsloth, reliability principles, and troubleshooting. It stresses that the training environment is ephemeral, so results must be pushed to the Hub or they are lost.

It targets ML engineers and researchers who want cloud fine-tuning without local infrastructure. Authentication uses an HF_TOKEN with write permissions passed through the job config's secrets mechanism, which is standard official Hugging Face tooling practice; the token references the user's own value and is used to push trained artifacts, not to exfiltrate data.