ml-research

Plugin ml-research from marketplace ml-plugins in krasserm/ml-plugins. Commands read from the repo's manifest — 1 month ago.

Skills (2)

• ml-research-loop

  Autonomous, budget-bounded loop that runs MANY ML experiments toward a goal with no human in the loop — hypothesize, train on HF Jobs, evaluate, keep/ discard, repeat until the target metric or budget is hit. Requires a `program.md` task spec (goal, keep/discard criteria, and a budget envelope). Use

• ml-research-task

  Autonomously research, write, and ship ML code on the Hugging Face ecosystem: training, fine-tuning (SFT/DPO/GRPO/LoRA), data processing, inference, and evaluation. Use whenever the task involves training or fine-tuning a model, building or auditing an HF dataset, running jobs on HF Jobs / cloud GPU

Subagents (2)

• ml-worker

  Runs exactly ONE ML experiment step end-to-end (submit a training job, or evaluate a finished one) and returns a compact result. Spawned by the ml-research-loop orchestrator per iteration to keep the campaign context lean. Not for direct use — the main agent uses the ml-research-task skill for one-o

• researcher

  Deep ML literature crawler. Mines papers for training recipes, validates datasets, and finds working code, returning a ranked recipe table. Delegate research-heavy lookups here (literature crawls, "what dataset/method/hparams produced result X", current TRL/Transformers API usage) to keep the main c

Hooks (1)

• PreToolUse

  uv run "${CLAUDE_PLUGIN_ROOT}/scripts/envelope_guard.py"