What's DVBench?¶

DVBench lets authors prove their models were trained on the data they state, and their performance on the test set are the ones that they report.

A DVBench entry consists of three layers of information:

• Attestation using a cryptographic tape named zkdv, you attest what checkpoints your model produced and what data it used during training.
• Measurements after a checkpoint is produced, you declare how and what you evaluate, and pre-register your evaluation code.
• Benchmarks when your model collected all measurements required for a leaderboard, it is automatically included.

This results in a trusted attestation of your entire training run, proving that it 1) used declared data during training and 2) measured the same checkpoint as the training produced.

Ready to jump in? Run your first training job.