List your company

Language Modeling on (val)

7.32PerplexityDAPT (nontoxic)
6.830810.132913.43516.7371Feb 8, 2022Oct 5, 2022Jun 1, 2023Jan 26, 2024Sep 21, 2024May 18, 2025Jan 13, 2026

Evaluation results

MethodLinks
DAPT (nontoxic)Parameter Size=530B2022.027.32———————
SGEAT (augmented)Parameter Size=530B, Training Samples=50k, Training Epochs=32022.027.86———————
SGEAT (augmented)Parameter Size=530B, Training Samples=100k, Training Epochs=32022.027.92———————
DAPT (nontoxic)Parameter Size=8.3B2022.028.12———————
SGEAT (augmented)Parameter Size=8.3B2022.028.91———————
SGEAT (augmented)Parameter Size=530B, Training Samples=50k, Training Epochs=52022.029.63———————
DAPT (nontoxic)Parameter Size=1.3B2022.0210.4———————
SGEAT (augmented)Parameter Size=1.3B2022.0211.19———————
DAPT (nontoxic)Parameter Size=357M2022.0213.33———————
Dynamic ExoFormerAnchor Source=External Anchor, Mixing Granularity=Elementwise, Normalization Strategy=Full2026.0114.09———————
ExoFormerAnchor Source=External Anchor, Mixing Granularity=Elementwise, Normalization Strategy=Full2026.0114.13———————
ExoFormerAnchor Source=External Anchor, Mixing Granularity=Headwise, Normalization Strategy=Full2026.0114.14———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Elementwise, Normalization Strategy=Full2026.0114.15———————
ExoFormerAnchor Source=External Anchor, Mixing Granularity=Scalar, Normalization Strategy=Full2026.0114.15———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Headwise, Normalization Strategy=Full2026.0114.17———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Scalar, Normalization Strategy=Full2026.0114.17———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Elementwise, Normalization Strategy=Only Q/K Norms2026.0114.21———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Headwise, Normalization Strategy=Only Q/K Norms2026.0114.21———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Scalar, Normalization Strategy=Only Q/K Norms2026.0114.22———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Headwise, Normalization Strategy=No Norm, Mixing Path Components={V,G}2026.0114.22———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Elementwise, Normalization Strategy=No Norm, Mixing Path Components={V,G}2026.0114.24———————
NuResFormerAnchor Source=Internal Anchor, Mixing Granularity=Scalar, Normalization Strategy=No Norm, Mixing Path Components={V,G}2026.0114.24———————
Naïve CombinationAnchor Source=Baselines2026.0114.25———————
ExoFormerAnchor Source=External Anchor, Mixing Granularity=Elementwise, Normalization Strategy=No Norm2026.0114.3———————
ResFormer (Value Residual)Anchor Source=Baselines2026.0114.32———————
SGEAT (augmented)Parameter Size=357M2022.0214.39———————
Gated AttentionAnchor Source=Baselines2026.0114.64———————
Base TransformerAnchor Source=Baselines2026.0114.79———————
DAPT (nontoxic)Parameter Size=126M2022.0217.97———————
SGEAT (augmented)Parameter Size=126M2022.0219.55———————
GDNlong-context fine-tuning=false, metric=perplexity2026.05—42.943.543.847.449.448.447.5
KLAlong-context fine-tuning=false, metric=perplexity2026.05—39.139.84043.645.344.543.6
Mamba2long-context fine-tuning=false, metric=perplexity2026.05—49.950.952.256.458.757.356.7