A living, friendly logbook of reproduction efforts. We run the paper, attach the harness, seeds, and hardware, and report clearly what we found — so you can see exactly how a result holds together, build on solid ground, and reuse the work yourself.https://frontiercheckpoint.com/https://frontiercheckpoint.com/reproductions/reproducing-nanogpt-speedrun/https://frontiercheckpoint.com/reproductions/reproducing-nanogpt-speedrun/The nanoGPT speedrun is a rare, fully open optimization target: hit 3.28 FineWeb validation loss on a GPT-2 (124M)-class model in minimum wall-clock on 8×H100. We reproduce the pipeline, isolate what the Muon optimizer and the architecture changes actually buy, and flag what will not transfer off the bench.Mon, 15 Jun 2026 00:00:00 GMTReproductionsreproducibilitypretrainingoptimizationdistributed-trainingeditors@frontiercheckpoint.com