Defines conventions for structuring, tracking, and reproducing machine learning experiments.
npx skills add https://github.com/MaxWolf-01/agents --skill ml-maxwolf-01
Please help me install this Agent Skill. Skill: ml Source: https://github.com/MaxWolf-01/agents/tree/main/mx/skills/ml Command: npx skills add https://github.com/MaxWolf-01/agents --skill ml-maxwolf-01
experiments/ with nested dataclass configs, where every result worth keeping becomes a Makefile target with frozen flags. - Mandatory run tracking: All runs log to wandb via a nested WandbConfig, with enable/disable handled at wandb.init rather than scattered conditionals. - Tensor code standards: einops/einx for rearrangements, runtime shape checking via torch-einops-utils or jaxtyping, and __main__ smoke checks in model files. - Hyperparameter tuning: Agent-assisted Optuna sweeps via autotune, which proposes search spaces from tyro scripts without YAML configs. - Use Case: When setting up a new training run, ask the agent to scaffold the experiment script, wire wandb tracking, and freeze the final invocation as a make target like figure-2a. ## Quick Start Set up a new ML experiment script with tyro config, wandb tracking, and a frozen make target for the baseline training run.High-intent search queries and answers about installing and using this skill.
Freeze each keeper result as a make target with its full invocation flags, such as `figure-2a: uv run python -u experiments/train_baseline.py --seed 42`. Reproduction then requires only hardware, a clone, `make install`, and the target name.
Define a nested `Config` dataclass per experiment, with shared pieces like `WandbConfig` and optimizer unions in the package. Expose everything you might sweep as a CLI flag via `tyro.cli`, never hardcoded.
Use einops for rearranges and reductions, with torch-einops-utils for runtime shape validation and no annotation machinery. jaxtyping plus beartype is the annotation-based alternative, requiring a ruff F722 ignore; pick one per project.
Only for a true smoke test, defined as something that could run in CI or as a pre-commit hook. Anything with real cycles or behavior you will inspect is a run and must be tracked, otherwise results evaporate.
Use `npx -y @roberttlange/autotune run experiments/<x>.py --trials N`, which proposes a search space from your tyro script, confirms it with you, and runs Optuna trials without modifying the file.
PaldomSet up Python virtual environments and manage Streamlit dependencies with uv.
yyccbbExtracts reproduction metrics with auditable scripts and compares them against paper tables and figures.
ArthTech-AIRun full-stack QA checks with pytest, Jest, ruff, TypeScript, and Playwright.
nishide-devGenerates Hydra configuration files for PyTorch Lightning ML training and Optuna sweeps.