List your company

Language Modeling on Domain-Specific Corpora Pretrain, Code, Law, Math (test)

87.82Pretrain ScoreLaw Dense Expert
21.426438.663255.973.1368Aug 15, 2024

Evaluation results

MethodLinks
Law Dense ExpertModel Architecture=Specialized, Number of Parameters=590M, Domain=Law2024.0887.8223.457.5330.7937.4
Math Dense ExpertModel Architecture=Specialized, Number of Parameters=590M, Domain=Math2024.0857.616.7810.435.6720.12
Code Dense ExpertModel Architecture=Specialized, Number of Parameters=590M, Domain=Code2024.0839.53.7110.838.7715.7
BTX (Baseline)Model Architecture=Generalist, Number of Parameters=590M2024.0826.723.786.635.7710.72
BAM CM (Shared KV)Model Architecture=Generalist, Number of Parameters=590M, KV Cache Strategy=Shared KV, Mixing Strategy=Coupled (CM)2024.08263.726.55.6310.46
BAM DM (Shared KV)Model Architecture=Generalist, Number of Parameters=590M, KV Cache Strategy=Shared KV, Mixing Strategy=Decoupled (DM)2024.0825.673.646.395.5710.32
BAM CM (Expert KV)Model Architecture=Generalist, Number of Parameters=590M, KV Cache Strategy=Expert KV, Mixing Strategy=Coupled (CM)2024.0824.843.536.215.4310
BAM DM (Expert KV)Model Architecture=Generalist, Number of Parameters=590M, KV Cache Strategy=Expert KV, Mixing Strategy=Decoupled (DM)2024.0824.023.456.045.319.7
Base ModelModel Architecture=Base, Number of Parameters=590M2024.0823.988.628.9214.1213.91