List your company

Language Modeling on FineWeb 100M token (val)

12.11PerplexitySpectron
11.537215.403619.2723.1364Feb 12, 2026

Evaluation results

MethodLinks
SpectronModel Architecture=Factorized Transformer-L, Number of Parameters=454M2026.0212.11——
Self guidedModel Architecture=Factorized Transformer-L, Number of Parameters=454M2026.0213.7——
Naive (AdamW)Model Architecture=Factorized Transformer-L, Number of Parameters=454M2026.0214.57——
SpectronModel Architecture=Factorized Transformer-M, Number of Parameters=297M2026.0214.62——
Self guidedModel Architecture=Factorized Transformer-M, Number of Parameters=297M2026.0215.53——
Naive (AdamW)Model Architecture=Factorized Transformer-M, Number of Parameters=297M2026.0215.54——
SpectronModel Architecture=Factorized Transformer-S, Number of Parameters=94M2026.0221.86——
Self guidedModel Architecture=Factorized Transformer-S, Number of Parameters=94M2026.0224.17——
Naive (AdamW)Model Architecture=Factorized Transformer-S, Number of Parameters=94M2026.0226.43——
BaselineTotal Epochs=2562026.06—3.047610
q0Total Epochs=2562026.06—3.003412
q0 (scaled)Total Epochs=9602026.06—2.98712.9