Agent Skill : LLM Training

Train and fine-tune language models using TRL (Transformer Reinforcement Learning) on Hugging Face Jobs infrastructure.

Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs package, UV scripts with PEP 723 format, dataset preparation and validation, hardware selection, cost estimation, Trackio monitoring, Hub authentication, and model persistence.

Skill Name: model-trainer

Version: 1.0.0

License: Apache-2.0

Main Documentation: skills/model-trainer/SKILL.md

Core Documentation

SKILL.md
skills/model-trainer/SKILL.md

Complete skill documentation with training guides, examples, and best practices

Training Scripts

Production-ready training scripts with Trackio integration, LoRA/PEFT, and Hub saving:

train_sft_example.py
skills/model-trainer/scripts/train_sft_example.py

Complete SFT (Supervised Fine-Tuning) training template with all best practices
train_dpo_example.py
skills/model-trainer/scripts/train_dpo_example.py

DPO (Direct Preference Optimization) training for preference learning
train_grpo_example.py
skills/model-trainer/scripts/train_grpo_example.py

GRPO (Group Relative Policy Optimization) training for online RL
convert_to_gguf.py
skills/model-trainer/scripts/convert_to_gguf.py

Complete GGUF conversion script for local deployment (Ollama, LM Studio, llama.cpp)
estimate_cost.py
skills/model-trainer/scripts/estimate_cost.py

Estimate training time and cost before launching jobs
dataset_inspector.py
skills/model-trainer/scripts/dataset_inspector.py

Validate dataset format before GPU training to prevent format failures

Reference Guides

Detailed guides for specific aspects of training:

training_methods.md
skills/model-trainer/references/training_methods.md

Overview of SFT, DPO, GRPO, KTO, PPO, and Reward Modeling methods
training_patterns.md
skills/model-trainer/references/training_patterns.md

Common training patterns and examples (quick demo, production, multi-GPU)
hardware_guide.md
skills/model-trainer/references/hardware_guide.md

Hardware specifications and selection guidelines for different model sizes
gguf_conversion.md
skills/model-trainer/references/gguf_conversion.md

Complete GGUF conversion guide with quantization options and usage examples
trackio_guide.md
skills/model-trainer/references/trackio_guide.md

Trackio monitoring setup for real-time training metrics visualization
hub_saving.md
skills/model-trainer/references/hub_saving.md

Hub authentication troubleshooting and model persistence best practices
reliability_principles.md
skills/model-trainer/references/reliability_principles.md

Reliability principles for production training workflows
troubleshooting.md
skills/model-trainer/references/troubleshooting.md

Common issues and solutions (OOM, timeouts, Hub push failures, etc.)

Plugin Configuration

plugin.json
plugin.json

Plugin metadata and configuration