Benchmarks

results.json holds every number on the model card; report.pdf is the full report. The evaluation archives are held out and not distributed.

run.py is the harness that produced the numbers. You can point it at your own archive (a folder with a tokens.u32 stream and a bank_valid.jsonl question bank in the same format as results.json describes) to run the same evaluation on your own data.