Lab · Records

Open a run record

A run record describes one training run: what was trained, on which data, with which settings, at what cost and with what result. The browser trainer writes one for every run. Open a record here to read it and to check that its specification matches the hash stored with it. A match shows only that the two agree. Anyone who edits the file can recompute the hash, so a match does not show who made the run, when, or that it ran as described.

Two examples are provided. The first is a 60-step run of the trainer’s Tiny model, made with the trainer’s code outside a browser. The second is eight training steps of nanochat, an open-source program for training small GPT-style models, on the text of 200 short addition problems. Its training loss is in nats per token and its validation loss in bits per byte, so the two are drawn on separate scales.