pi-ml-intern

ML Intern extension for Pi — autonomous ML research & implementation with literature-backed recipes. Heavily inspired by Hugging Face's ml-intern.

Only activates when you explicitly invoke /ml-intern, with no impact on Pi's default behavior.

⚠️ EXPERIMENTAL — This extension is under active development (v0.2.0). Performance may lag behind the original project. Feedback and contributions welcome.

⚠️ CAUTION — By default, hf_jobs sets visibility as public for models trained in Hugging Face.

Quick Start

Install

pi install npm:@santiago-r/pi-ml-intern

Or manually: copy this directory to ~/.pi/agent/extensions/ml-intern/ (global) or .pi/extensions/ml-intern/ (project-local).

Usage example

/ml-intern Generate a state-of-the-art cardiac event classifier (multi-label) from publicly available ECG datasets. Emphasize optimization of low-level signal pre-processing. Emphasize explainability, output SHAP values alongside predictions. Evaluate on arrythmia detection ROC-AUC for a held-out test split.

Type /ml-intern followed by your ML task. Include concrete and relevant details, instructions, constraints, and evaluation criteria.

Requirements

• Pi (coding agent harness)

Recommended access tokens

Set these for full capability. Without them, HF Hub lookups and GitHub code searches are rate-limited to public repos only, and HF Jobs is unavailable.

• HF_TOKEN — enables gated/private datasets, model downloads, and HF Jobs GPU training. Get one here.
• GITHUB_TOKEN — lifts GitHub API rate limits. Get one here.

Tokens are auto-loaded from .env in the working directory. No manual export needed.

What it does

When you use /ml-intern, the agent:

1. Researches first — finds landmark papers, crawls citation graphs, reads methodology sections
2. Validates resources — checks dataset schemas, model architectures, API compatibility
3. Implements — includes instructions against hallucinated imports, wrong arguments, and silent substitutions
4. Tracks progress — uses plan_tool for tasks with 3+ steps

Tools included

| Tool | Description | | --- | --- | | plan_tool | Track ML task progress (pending/in_progress/completed) | | hf_papers | Paper search, citation graphs, methodology section reading | | hf_jobs | Submit & manage remote GPU/CPU compute jobs on HF Cloud | | hub_repo_details | HF Hub model/dataset/space details and search | | hf_inspect_dataset | Dataset schema, splits, and sample inspection | | github_find_examples | Find working ML example scripts in GitHub repos | | github_list_repos | Discover repos for GitHub orgs/users | | github_read_file | Read file contents from GitHub repos | | explore_hf_docs | Browse HF documentation structure | | fetch_hf_docs | Fetch full HF documentation pages | | find_hf_api | Search HF REST API endpoints | | research | Spawn isolated sub-agent for deep literature research |

Running headless (no TUI)

Prefer the CLI? Set ML_INTERN_FORCE=1 for print mode:

ML_INTERN_FORCE=1 pi -p "Train a small GPT on input_data/my_data.jsonl"

This activates all 12 research tools and the ml-intern system prompt for the duration of the command, then exits. Works in shell scripts, CI, cron jobs, or even in agent subtasks if you are crazy enough.

Attribution

Heavily inspired by Hugging Face's ml-intern (Apache 2.0). All original TypeScript code is Apache 2.0-licensed.

License

Apache 2.0 — see LICENSE.