ml-intern

Audit an HF dataset — schema, sample rows, anomalies, recommended training method.

Kick off the full ml-intern workflow on an ML task — research → audit dataset → architect training job → submit. Loads the ml-intern skill and dispatches the right subagents.

Pre-flight a training script before submitting it to HF Jobs — checks for the 8 expensive mistakes.

Deep literature crawl. 6–10 query angles, 2-hop citation graph BFS, 30–50 full-paper reads in parallel subagents, cross-paper synthesis with gap analysis.

Run a literature review for an ML task — finds landmark paper, crawls citation graph, extracts recipe.

Dataset quality auditor for HF datasets. Use before committing to a dataset for fine-tuning. Returns schema, row counts, sample rows, distributions, anomalies (class imbalance, duplicates, missing values, format issues), and a recommended training method based on column shape. Isolates 10k+ tokens of dataset metadata + sample rows from the main thread.

Single-paper deep reader. Reads ONE paper end-to-end (abstract → intro → method → experiments → results → limitations → future work) and returns a structured ~800-word digest where every factual claim is backed by a verbatim quote with §section reference. Designed for parallel fan-out from `/ml-research-ultra` — each invocation isolates 50k+ tokens of paper HTML from the main thread. Use when the orchestrator needs the full content of a paper, not just the recipe.

ML literature crawler. Use when the main task needs a methodology-grounded recipe drawn from multiple papers — e.g., "find the best recipe for math reasoning fine-tuning", "what dataset and method does the GRPO follow-up work use", "literature review for sparse-attention long-context training". Returns a structured ≤800-word report with anchor papers, extracted recipes, citation-graph descendants, and working code-example URLs. Isolates 50k+ tokens of paper text from the main thread.

Designs and reviews ML training submissions for both local execution and HF Jobs. Use after the recipe is chosen and the dataset is audited — produces a complete training script + the exact run command, sized to hardware, with all required fields (push_to_hub, hub_model_id, disable_tqdm, Trackio, timeout, package installs). Detects compute mode automatically and asks the user when both local and Jobs are viable. Catches the "model lost" / "30m timeout" / "missing flash-attn" mistakes before they cost real money.

Use when the user asks to fine-tune, train, evaluate, audit, or ship a machine-learning model on the Hugging Face ecosystem — SFT, DPO, GRPO, RLHF, LoRA/QLoRA, post-training, dataset auditing, paper-driven research, hf jobs submission, Trackio monitoring, push-to-Hub. Triggers include "fine-tune", "train a model", "SFT", "DPO", "GRPO", "RLHF", "post-training", "audit this dataset", "literature review for X task", "submit hf job", "find a dataset for X", "best recipe for X", "hyperparameter sweep", "OOM during training", "push to Hub". Replicates the workflow of huggingface/ml-intern inside Claude Code with zero new dependencies.

Use when the user names a specific ML model (e.g. DINOv3, SAM 2, Whisper, Qwen2-VL) and wants its real/official code, training recipe, or papers found, verified, or archived locally for grounded coding. Triggers include "find the real code for this model", "harvest DINOv3", "store the model's code and papers locally", "archive the training/inference code", "set up a local source-of-truth / reference archive for a model", or any request that future coding against a model be grounded in its actual source instead of training-time memory.