training-hub
Fine-tune LLMs using Red Hat training-hub library with SFT, LoRA, and OSFT algorithms. Use when prep
作者 RobbieJ|オープンソース
Training Hub Skills for Claude Code
A Claude Code skill for fine-tuning LLMs using Red Hat training-hub.
Features
- Three training algorithms: SFT, LoRA/QLoRA, and OSFT (continual learning)
- Hardware profiles: Optimized configs for RTX 4090/5090, DGX Spark, A100, H100, H200
- Scale support: Single GPU to multi-node datacenter clusters (Slurm, Kubernetes)
- Utility scripts: Memory estimation, config recommendation, dataset validation
- Full workflow: Dataset prep → Training → Evaluation → Inference serving
Installation
As a Claude Code Plugin (Recommended)
claude plugins add github:robbiej/training-hub-skills
Manual Installation
Clone to your personal skills directory:
git clone https://github.com/robbiej/training-hub-skills ~/.claude/skills/llm-training
Or for a specific project:
git clone https://github.com/robbiej/training-hub-skills .claude/skills/llm-training
Quick Start
Once installed, Claude Code automatically uses this skill when you ask about:
- Fine-tuning LLMs with training-hub
- Preparing JSONL datasets
- Configuring training for specific hardware
- Scaling to multi-GPU or multi-node setups
- Evaluating models with lm-eval
- Serving models with vLLM
Example Prompts
"Help me fine-tune Llama 3.1 8B on my RTX 5090"
"What's the best config for training a 70B model on 8 H100s?"
"Convert my CSV dataset to JSONL for training-hub"
"How do I evaluate my fine-tuned model?"
Utility Scripts
# Get optimal config for your hardware
python scripts/recommend_config.py --model meta-llama/Llama-3.1-8B-Instruct --hardware rtx-5090
# Check memory requirements
python scripts/estimate_memory.py --model meta-llama/Llama-3.1-70B-Instruct --method lora --hardware h100
# Validate dataset before training
python scripts/validate_dataset.py ./training_data.jsonl
# Convert from other formats
python scripts/convert_to_jsonl.py alpaca input.json output.jsonl
Documentation
| File | Description |
|---|---|
| SKILL.md | Main skill entry point |
| HARDWARE.md | Hardware-specific configurations |
| ALGORITHMS.md | SFT, LoRA, OSFT parameter reference |
| DATA-FORMATS.md | JSONL format and conversion |
| SCALE.md | Multi-node, Slurm, Kubernetes |
| EVALUATION.md | lm-evaluation-harness integration |
| INFERENCE.md | vLLM serving and deployment |
| ADVANCED.md | Gradient checkpointing, flash attention |
| MODELS.md | Pre-tested model configurations |
| TROUBLESHOOTING.md | Common issues and solutions |
Supported Hardware
| Hardware | VRAM | Best For |
|---|---|---|
| RTX 4090 | 24GB | 7B LoRA, 34B QLoRA |
| RTX 5090 | 32GB | 8B LoRA, 70B QLoRA |
| DGX Spark | 128GB | 70B SFT |
| A100 80GB | 80GB | 14B SFT, 70B LoRA |
| H100 80GB | 80GB | 14B SFT, 70B LoRA |
| H200 141GB | 141GB | 32B SFT, 140B LoRA |
| 8×H100 | 640GB | 70B SFT |
| 8×H200 | 1.1TB | 140B+ SFT |
Requirements
- training-hub for training
- vLLM for inference (optional)
- lm-eval for evaluation (optional)
License
Apache 2.0
Contributing
Contributions welcome! Please open issues or PRs on GitHub.