GoodSeed

ML experiment tracker. Logs metrics and configs to local SQLite files, serves them via a built-in HTTP server, and visualizes them in the browser.

Full documentation at docs.goodseed.ai.

Install

pip install goodseed

Python 3.9+ required.

For development:

uv sync --extra dev

Quick Start

Log metrics and configs from a training script:

import goodseed

run = goodseed.Run(name="my-experiment", tags=["bert"])

# Log configs (single values)
run["learning_rate"] = 0.001
run["batch_size"] = 32

# Log metrics (series of values)
for epoch in range(100):
    loss = train_step()
    run["train/loss"].log(loss, step=epoch)

run.close()

Custom step values:

run["metric"].log(value=acc, step=i)

step is required for series logging via run["path"].log(...).

Batch logging methods are also available:

run.log_configs({"learning_rate": 0.001, "batch_size": 32})
run.log_metrics({"loss": loss, "acc": acc}, step=step)

Your data is saved to a local SQLite file. You can also use with goodseed.Run(...) as run: to close the run automatically. Goodseed also registers best-effort per-run atexit cleanup, but calling run.close() (or using with) is still the recommended way to guarantee upload completion.

Storage Modes

The storage parameter controls where data is stored:

• "cloud" (default) — local SQLite plus background sync to the remote API.
• "local" — local SQLite only, no remote sync.
• "disabled" — no storage; all writes are silent no-ops.

Cloud storage syncs data in a background thread while your training runs. Set the GOODSEED_API_KEY environment variable and use workspace/project format for the project name:

import goodseed

run = goodseed.Run(project="my-workspace/my-project", name="experiment-1")
run["train/loss"].log(0.5, step=0)
run.close()  # blocks until all data is uploaded

For local-only storage (no remote sync):

run = goodseed.Run(storage="local")

You can also set the mode via the GOODSEED_STORAGE environment variable.

Read Data from Server

run = goodseed.Run(project="my-workspace/my-project", run_id="bold-falcon", read_only=True)
data = run.get_metric_data("train/loss")
configs = run.get_configs()

Resume a run

run = goodseed.Run(resume_run_id="bold-falcon")
run["train/loss"].log(0.3, step=123)
run["eval/f1"] = 0.85
run.close()

Monitoring

By default, goodseed automatically captures:

• stdout / stderr — every print() and warning is logged
• Tracebacks — captured on unhandled exceptions (status set to failed)
• CPU & memory — via psutil (installed with goodseed)
• GPU — NVIDIA (via nvidia-smi) and AMD (via rocm-smi), no pip dependency

Disable any of these with:

run = goodseed.Run(
    capture_stdout=False,
    capture_stderr=False,
    capture_hardware_metrics=False,
    capture_traceback=False,
)

Then view your runs:

goodseed serve

Open the printed link in your browser to see your runs, metrics, and configs.

Configuration

Git Tracking

By default, Run() auto-tracks Git metadata from the current repository:

• dirty state
• source_code/diff (index vs HEAD)
• last commit message, ID, author, and date
• current branch
• remotes
• source_code/diff_upstream_<sha> when HEAD differs from the remote tracking branch

Specify a custom repository path:

import goodseed

run = goodseed.Run(git_ref=goodseed.GitRef(repository_path="/path/to/repo"))

Disable Git tracking:

import goodseed

run = goodseed.Run(git_ref=False)
# or: run = goodseed.Run(git_ref=goodseed.GitRef.DISABLED)

CLI

goodseed                   # Start the server (default command)
goodseed serve [dir]       # Start the server, optionally from a specific directory
goodseed serve --port 9000 # Use a custom port
goodseed list              # List projects
goodseed list -p default   # List runs in a project
goodseed upload -p <workspace/project> --run-id <run_id>  # Upload one run
goodseed upload -p <workspace/project>                    # Upload all runs

Tests

uv sync --extra dev
uv run pytest tests/ -v

See DOCS.md for architecture details and API reference.

Beta Notice

GoodSeed is currently in beta. We may introduce breaking changes as we iterate on the product.

In particular, the local SQLite schema and parts of the Python/CLI interface may change in future releases. Depending on the change, upgrading could require migrating or recreating existing local GoodSeed data files (stored under ~/.goodseed by default).

Feedback is very welcome while we stabilize the API. Please open issues with bug reports or feature requests.