weightslab

Hooks into a live PyTorch training loop to surface mislabels, outliers, and class imbalance without restarting.

Pluginfreeglobal

WeightsLab is a free, open-source tool that hooks into a live PyTorch training loop, letting ML engineers pause training, sort samples by loss to find mislabels and outliers, relabel or drop them, and resume from the same checkpoint without restarting. It's built for messy real-world image, video, and LiDAR data, connecting training signals back to the exact samples causing them rather than just logging aggregate metrics like a typical experiment tracker. It's aimed at AI engineers who want to fix their data live during training instead of debugging it after the fact.

Website

Full match profile

Behind the summary, Matchbox keeps a richer profile of weightslab - the signals our matcher actually reads to decide when to surface it. It stays private; claim the listing to see and control it.

Problem & pain-point mapping
Who we surface it to (audience fit)
What it's a strong alternative to
Trust & credibility signals

Claim weightslab to unlock →

Try Matchbox with your own problem

Describe what is not working - we’ll show you whether weightslab (or something else) actually fits.

Try a problem