You can deploy this model and start earning money today!

Hoff1610/
ml-intern-claude-code

Views

No views yet

🤖 ML Intern for Claude Code

An autonomous ML engineering agent that brings the full power of the Hugging Face ecosystem into your Claude Code environment. It can train models, manage datasets, research papers, run evaluations, and ship production ML — all through natural language.

What You Get

Capability	How
🏋️ Model Training	SFT, DPO, GRPO, Reward Modeling via TRL on HF Jobs (cloud GPUs)
📊 Dataset Management	Search, validate, inspect, filter datasets on HF Hub
🔬 Paper Research	Search, read, and extract training recipes from ML papers
📈 Experiment Tracking	Real-time Trackio dashboards with alerts
🏆 Model Evaluation	inspect-ai and lighteval on local or cloud hardware
🔍 Model Discovery	Benchmark-based model selection and comparison
👁️ Vision Training	Object detection, classification, SAM segmentation
🌐 Gradio Apps	Build web UIs and demos
🔧 Hub Operations	Full repo management, webhooks, collections, endpoints

Quick Setup

bash

# 1. Clone this repo
git clone https://huggingface.co/Hoff1610/ml-intern-claude-code
cd ml-intern-claude-code

# 2. Set your HF token
export HF_TOKEN=hf_your_token_here

# 3. Run setup
chmod +x setup.sh
./setup.sh

# 4. Open Claude Code in your project
claude

The setup script will:

Install the hf CLI and uv
Copy CLAUDE.md (system prompt) and .claude/ config to your project
Install HF Skills (LLM trainer, datasets, trackio, evals, papers)
Configure the HF MCP Server for Claude Code
Set up safety hooks (block dangerous commands, audit logging)

Manual Setup

1. Copy files to your project

your-project/
├── CLAUDE.md                          ← copy from this repo
└── .claude/
    ├── settings.json                  ← copy from this repo
    └── hooks/
        ├── pre_bash.sh                ← copy + chmod +x
        ├── post_tool_log.sh           ← copy + chmod +x
        └── on_stop.sh                 ← copy + chmod +x

2. Add HF MCP Server

claude mcp add huggingface -t http "https://huggingface.co/mcp?login"

3. Install HF Skills

bash

pip install -U "huggingface_hub[cli]"
hf skills add --claude --global

4. Enable MCP tools

Go to huggingface.co/settings/mcp and enable all tools.

File Structure

ml-intern-claude-code/
├── README.md                    ← This file
├── CLAUDE.md                    ← System prompt with ML expertise + workflows + guardrails
├── setup.sh                     ← One-command installer
└── .claude/
    ├── settings.json            ← MCP servers (HF + GitHub) + hooks config
    └── hooks/
        ├── pre_bash.sh          ← Safety: blocks destructive commands + token exposure
        ├── post_tool_log.sh     ← Audit: logs all tool calls to JSONL
        └── on_stop.sh           ← Status: reports running HF Jobs on exit

Usage Examples

> Train Qwen2.5-0.5B on the Capybara dataset with SFT
> What's the best model for code generation under 7B params?
> Find the latest papers on RLHF and summarize the top approaches
> Evaluate meta-llama/Llama-3.2-1B on MMLU
> Inspect trl-lib/ultrafeedback_binarized for DPO compatibility

Architecture

Claude Code
    ├── CLAUDE.md (system prompt with ML expertise)
    ├── HF MCP Server → model/dataset/paper search, jobs, docs
    ├── HF Skills → training, tracking, evals, datasets, papers
    ├── Hooks → safety, audit logging, job status
    └── GitHub MCP Server (optional) → code search

Requirements

Requirement	Why
Claude Code	Agent runtime
HF Token (write access)	Hub auth + Jobs + MCP
HF Pro/Team/Enterprise	Required for HF Jobs
`jq`	Hook JSON parsing
`uv`	PEP 723 training scripts

License

MIT