trackle-ml 0.1.1

Local-first, git-aware ML experiment tracking.

Trackle

Trackle is a lightweight, local-first, git-aware ML experiment tracker.

Goals

• Local filesystem storage only
• No servers or networking
• Simple Python API and CLI
• Git-based reproducibility

Quick start

pip install -e .

import trackle

trackle.init(run_name="demo")
trackle.log_params({"lr": 0.01})
trackle.log_metric("loss", 0.5, step=1)
trackle.finish()

Storage layout

Runs are stored locally (default ./trackle_experiments):

trackle_experiments/
  runs/
    <run_id>/
      run.json       # metadata: name, created_at, finished_at, duration_s, tags
      params.json    # merged params
      metrics.csv    # step,key,value,timestamp
      context.json   # env info (python/os/hw/torch)
      git.json       # git commit/branch/dirty/remote
      artifacts/     # copied artifacts
      notes.md       # optional note

CLI

trackle list-runs [--base DIR] [--limit N]
trackle diff-runs <run_a> <run_b> [--base DIR]
trackle plot-metrics <run_id> [--key metric] [--base DIR]

Extras

• pip install trackle[viz] to enable matplotlib plotting in the CLI.
• pip install trackle[rich] to enable richer CLI output (if extended later).