dictlearn

dictlearn

Dictionary learning next-token accuracy evaluation.

Feeds each val sequence through the model and checks whether the argmax prediction at the SEP position equals the correct result token.

See theseus.data.datasets.dictlearn for the full dataset description.

Registered variants mirror the dataset registry: - dictlearn_16, dictlearn_512 (default, 64 values) - dictlearn_{16,512}_v{32,64,128,256,512,1024}