Dohnuts documentation

Use Dohnuts

• Installation and inference: load a checkpoint and make text or image decisions.

• Agent integration: expose decisions through the Bub SDK.

• Training and evaluation: run the complete workflow or resume training.

Understand the model

• Model card: capabilities, training, measured results, and limitations.

• Benchmark comparisons: Jev, Laya multilingual, and Laya Vision results.

• Architecture: candidate scoring, shared computation, and model adapters.

• RLCD: objective, sampling, and calibration.

• Compute efficiency: fused execution and the shared training/inference path.

Reproduce the measurements

• Data and evaluation: datasets, partitions, and quality metrics.

• Reference protocols: Laya and Laya Vision checkpoints and comparison scope.

• Benchmarking: latency, resource measurements, JevBench, and figure generation.

The behavior and regression tests cover the prediction interface and known failure cases.

For contributors, see Development and Build the website.