This dataset captures per-step training and validation metrics from training runs of a 12-layer GPT-style decoder-only transformer. Each run is stored as a single .csv file in which every row corresponds to one logged step, and several columns hold per-parameter measurements encoded as JSON.
The dataset is designed to support post-hoc analysis of:
