Inspect — probe & train the checkpoint

A trained checkpoint is not a black box. This page opens the model your Console session trained and lets you look inside it three ways:

• Probe a sentence — a read-only forward pass over any sentence you type: per-head attention maps at the layer you choose, Q/K/V tensor statistics, and the model's top-k prediction for every position.
• Live training iterator — start a fresh, untrained copy of the model and train it one optimiser step at a time. The optimiser state lives in your session on the server, so "Next iteration" genuinely continues the same training run — watch the loss curve fall as the model memorises its batch.
• Step-by-step batch trace — one complete training step (forward → loss → backward → update) executed on a copy of your trained checkpoint, with the gradient and weight-delta norms that prove an update happened. Your session's model is never modified.

Checking access…