Edit

Two local JAX training runs for attestation and dataset-filter checks.

tape_validation/20260905_201353/next_token_mse vs. estimated ISO FLOPs

2 matching runs · 6ND; missing tokens estimated at 20× parameters
All 2 filtered runs
Run Scale (B) ISO FLOPs Training datatape_validation/20260905_201353/next_token_mse
tape_validation/20260905_201353_baking_window8
No description yet
<0.001~3.8e30.04
tape_validation/20260905_201353_astronomy_window8
No description yet
<0.001~1.9e30.07
2 models