Pipeline console & Atlas (live BERT)

Every track before this one built a single component at demonstration scale — a handful of tokens, matrices small enough to check by eye. This track connects the components and turns the machine on: a real PyTorch BERT, small enough to train from scratch on your own corpus, right here. Which sets up the experiment the whole pillar has been building toward: at the start of training, the model's guess for a masked word is pure noise. A few hundred steps later it is guessing your words back at you. Can you catch the moment it starts being right?

Three modules split the work. The Atlas comes first and runs nothing at all: it lays the eight stages out as one connected map, so that tokenisation, embeddings, attention and the encoder stack stop being separate chapters and become stations on a single route. The Console then drives that route end to end — corpus in, model trained, and the numbers at each step opened up: real token ids, real query-key-value matrices, real attention weights. Inspect goes below deck: probe the trained checkpoint's heads layer by layer, watch the loss fall and the weight matrices drift as you keep training, and save the result as your own checkpoint.

This is also where the earlier tracks' promises get audited. The attention pass you once checked by hand shows up as heatmaps computed at full speed; the encoder block you assembled runs as a stack inside a live model; and the claim underneath everything — that sentences meaning similar things end up with vectors sitting near each other — stops being something you were told and becomes something you can test against a model you trained yourself.

The map below is the route in one picture; the modules walk it station by station.