Hunch

Every hunch, tracked.

A journal for your ML experiments. Every training run saves its code, config, environment and metrics, why you ran it, and which run it built on.

pip install sigmoid-hunch

If you know git, you know Hunch

Why not just git? Twenty runs of the same code with different learning rates are one commit to git. Hunch keeps all twenty, each with its settings and results.

In your code

hunch run already saves the code, command and environment. Add two lines and Hunch also tracks your settings and metrics, so you can see what changed between runs and how each one trained:

import hunch

hunch.init(config=args)                           # your settings: argparse args, a dict or a dataclass
for step in range(steps):
    ...
    hunch.log({"val_loss": val_loss}, step=step)  # your metrics, drawn as curves

Running python train.py without hunch run? hunch.init() records the run on its own; add why="..." to say why. Same names as W&B: import hunch as wandb works.

In your editor

Hunch for PyCharm: your runs as a tree, keep or drop them, compare any two in the diff viewer, and start runs with one click. In review on JetBrains Marketplace.

VS Code, Cursor and Windsurf: an extension is on the way.

Your disk, your files

No server, no account, no telemetry. Just files in .hunch/.

Guide · File format · Questions · Contributing

Share it

Saved you a "what did I run last month?" moment? Tell one person, or post your hunch log in Show and tell.

Tracking a research repo with Hunch? Show it:

[![tracked with Hunch](https://img.shields.io/badge/tracked_with-hunch-6E56CF?style=flat-square)](https://github.com/sigmoid-ide/hunch)

Apache-2.0 · Part of Sigmoid IDE · Like it? Give it a star.