model-trainer

What problem does it solve?

This skill automates cloud-based TRL training workflows on Hugging Face Jobs, enabling SFT, DPO, GRPO and reward modeling training without local GPU infrastructure, plus GGUF conversion for local deployment.

Core Features & Use Cases

• Cloud TRL training: Run TRL methods (SFT, DPO, GRPO, Reward Modeling) on Hugging Face Jobs infrastructure.
• GGUF conversion: Convert trained models to GGUF format for local inference with Ollama, llama.cpp, etc.
• End-to-end workflow: Include dataset preparation/validation, hardware selection, hub authentication, Trackio monitoring, and cost estimation.
• Guided execution: Use hf_jobs() MCP tool to submit training scripts; Trackio for real-time monitoring; Hub push to persist results.

Quick Start

Submit a complete training job via hf_jobs() using the template script (e.g., scripts/train_sft_example.py) for your model, dataset, and preferred hardware. The job will run on cloud GPUs, monitor via Trackio, and push results to the Hub.