ml-intern
- Type
- repo
- Venue
- Hugging Face (GitHub)
- Year
- 2026
- Source
- github
- Access
- free
- Added
- 2026-08-14T18:45:00Z
- Verified
- 2026-08-14T18:45:00Z
Summary
huggingface/ml-intern is a CLI/agent (smolagents lineage) that reads HF docs/papers/datasets, writes training code, and can run bash/read/write/edit locally or opt into HF Space sandbox tools and HF Jobs. Hosted inference goes through HF Inference Providers (HF_TOKEN); local models via LiteLLM prefixes ollama/vllm/lm_studio/llamacpp. Sessions auto-upload as Claude-Code-style JSONL traces to a private {user}/ml-intern-sessions Hub dataset (HF Agent Trace Viewer). Slack one-way notifications. Agentic loop max 300 iterations with 170k auto-compaction and a doom-loop detector. Apache-2.0. Python >=3.11. pip/uv tool install from the repo.
Keywords
ml-intern · huggingface · coding-agent · smolagents · traces · cli · hf-jobs
Topics
ML engineering agents, coding agents
Research notes
- Primary: GitHub README + API (Apache-2.0, Python, 10726 stars / 1169 forks at check). Cite authors from README bibtex. Demo space https://smolagents-ml-intern.hf.space/. Related public dataset clem/ml-intern-sessions is already papers_export id 380; leftover is the GitHub repo so hf_* left blank. Discord posted the repo. Per-user private Hub traces are not a new public hosted corpus, so no datasets_local row.