ultra-ml-intern: ML engineering intern for Claude Code

ultra-instinct ML engineering intern for Claude Code. Reads papers, audits datasets, ships SFT/DPO/LoRA runs to Hugging Face.

ultra-ml-intern is a Claude Code plugin that gives Claude the workflow of an ML engineering intern. It researches ML papers, audits Hugging Face datasets, designs fine-tuning recipes (SFT, DPO, GRPO, LoRA, QLoRA, RLHF), and submits training jobs to HF Jobs with Trackio monitoring.

The procedural knowledge comes from huggingface/ml-intern, HF's standalone Python harness around the Claude API. This repo wires the same intelligence into Claude Code, Anthropic's official agentic harness for Claude. Same model, a more capable loop, and you bring your own Claude (Max subscription or API key) instead of paying for a second harness on top.

Works in any Claude Code surface: terminal CLI, IDE extensions, and the web app.

Install

# In any Claude Code session:
/plugin marketplace add infiniV/ultra-ml-intern
/plugin install ml-intern@ultra-ml-intern

Restart Claude Code, then verify with /plugin and /agents. The slash commands (/ml-intern, /ml-research, …) keep their short names; the ultra- prefix is just the package wrapper.

What you get:

• 1 skill: ml-intern (the workflow)
• 5 slash commands: /ml-intern, /ml-research, /ml-audit, /ml-preflight, /ml-train
• 3 subagents: ml-paper-researcher, dataset-auditor, training-job-architect
• 1 MCP server: Hugging Face (activates when HF_TOKEN is set)

Quickstart

> "fine-tune Qwen3-0.5B for math reasoning"

The skill activates automatically and walks the 6-step research-driven workflow:

1. Find the landmark paper for the task
2. Crawl the citation graph for recent SOTA
3. Read methodology sections (3, 4, 5) and extract the recipe
4. Validate the dataset and base model exist on Hub
5. Write a training script grounded in current TRL APIs
6. Pre-flight check → smoke test → full hf jobs run with Trackio monitoring

What it does

You ask	It does
"fine-tune X for Y"	Full pipeline: literature review → dataset audit → training-job design → smoke test → full run
"what's the best recipe for X"	Dispatches the ml-paper-researcher subagent; returns recipe + citations
"audit dataset Y"	Dispatches the dataset-auditor; returns schema, anomalies, GO/NO-GO verdict
"preflight train.py"	Catches missing push_to_hub, default 30m timeout, bf16 on T4, missing flash-attn install, before you spend cluster hours
"submit hf jobs run"	Walks pre-flight → cost estimate → smoke test → full submission → Trackio dashboard URL

Commands

Command	What it does
/ml-intern	The full pipeline: research → audit → train → ship
/ml-research	Literature review only: landmark paper, citation graph, extracted recipe
/ml-audit	Dataset audit only: schema, samples, anomalies, training-method recommendation
/ml-preflight	Sanity-check a training script before submission
/ml-train	Submit a training job. Local-first when a GPU is available, HF Jobs when not

Subagents

Subagent	Role
ml-paper-researcher	Crawls arXiv + cites the landmark paper, extracts the methodology section into a recipe
dataset-auditor	Inspects HF datasets: schema, sample rows, distribution checks, anomaly flagging
training-job-architect	Writes the TRL/Transformers training script + the hf jobs run command sized to your hardware

Training recipes supported

The plugin recognizes and writes scripts for:

• SFT (Supervised Fine-Tuning) for single-turn and multi-turn chat
• DPO (Direct Preference Optimization) on pairwise preference data
• GRPO (Group Relative Policy Optimization) for reasoning tasks, DeepSeek-style
• LoRA and QLoRA for parameter-efficient fine-tuning, including 4-bit quantization
• RLHF with the full reward-model plus PPO pipeline
• Continued pretraining for domain adaptation on raw text

All grounded in the current TRL API. The ml-paper-researcher reads the actual library source, not its training data, before writing imports.

Hardware sizing built in

The plugin knows the HF Jobs flavors (t4-small through a100x8) and picks one to fit your model:

Model size	Default flavor
1–3B	a10g-largex2 (48 GB)
7–13B	a100-large (80 GB)
30B+	l40sx4 or a100x4
70B+	a100x8

Full chart in references/hardware-sizing.md.

Mistakes it prevents

The 8 expensive errors from huggingface/ml-intern's system prompt, encoded here as procedural rules:

1. Hallucinated TRL/Transformers imports → plugin reads the current source first
2. Wrong trainer arguments → fetches the actual config docs before writing
3. Wrong dataset format → inspect_dataset.sh runs first
4. Default 30m timeout kills the job → minimum 2h enforced
5. Lost models (no push_to_hub=True) → preflight refuses without it
6. Batch submission failures → smoke-test one job first
7. Silent dataset substitution → surfaces the failure and asks the user
8. Hardcoded missing packages (flash-attn, etc.) → preflight catches it

One more rule worth calling out: no scope-changing fixes. OOM doesn't mean silently rewriting SFT into LoRA. It means a proper batch and grad-accum reduction.

Requirements

• Claude Code (any surface: terminal, IDE, web)
• Your own Claude access (Max subscription or API key) for the model itself
• Bash + standard Unix tools
• For HF Jobs submission: a Hugging Face account with billing enabled
• Optional: HF_TOKEN exported for the bundled MCP server

export HF_TOKEN="$(hf auth print-token)"  # or paste from https://huggingface.co/settings/tokens

The MCP server adds Hub doc semantic search and community Gradio Space tools. The plugin works without it. It falls back to WebFetch and the bundled inspect_dataset.sh, crawl_arxiv.sh, and hf_paper_meta.sh helpers.

Relationship to huggingface/ml-intern

Both projects use the same model, Claude. What differs is the harness around it.

The upstream is a standalone Python project: roughly 50k lines of agent loop, plan tool, paper and dataset tools, jobs tool, and a hand-built system prompt. It calls the Claude API directly. It shipped first and the procedural knowledge in this repo is theirs.

Claude Code is Anthropic's official agentic harness for Claude. The agent loop, the planner (TodoWrite), subagent dispatch, MCP plumbing, and the editor surface are already there and tested. This plugin wires the procedural knowledge from the upstream into those primitives, so Claude runs the same workflow without a parallel harness sitting on top of the API.

In practice: same Claude model, a more capable surrounding loop, no extra billing layer (your Claude Max subscription or API key is enough), and the workflow runs anywhere Claude Code already runs.

If you want the standalone Python tool, use the upstream. If you already work in Claude Code, use this.

Uninstall

/plugin uninstall ml-intern@ultra-ml-intern
/plugin marketplace remove ultra-ml-intern

Contributing

Issues and PRs welcome at https://github.com/infiniV/ultra-ml-intern.

CI runs:

• shellcheck on every script in skills/ml-intern/scripts/
• bash -n syntax check on the same
• JSON schema validation on .claude-plugin/plugin.json and .claude-plugin/marketplace.json

Credits

• Upstream: huggingface/ml-intern by the Hugging Face team. The procedural knowledge is theirs.
• Plugin format reference: obra/superpowers.
• Plugin skeleton mirrors infiniV/claude-code-audio-notify.

License

MIT