List your company

Language Modeling on WikiText-103 byte-level (val)

1.796Loss (Context 8192)RoPE+ALiBi
1.764721.975862.1872.39814May 5, 2026

Evaluation results

MethodLinks
RoPE+ALiBiEvaluation context length=81922026.051.7960.028—
Scaled-exact Jordan-RoPEEvaluation context length=8192, c=12026.051.8690.0310.0014
Damped RoPEEvaluation context length=81922026.051.8840.0510
Scaled-exact Jordan-RoPEEvaluation context length=8192, c=.1, large eta init=true2026.051.8990.0340.0223
Stabilized Jordan-RoPEEvaluation context length=81922026.051.9030.0590.0013
Scaled-exact Jordan-RoPEEvaluation context length=8192, c=.12026.051.9070.0470.0021
ALiBiEvaluation context length=81922026.051.9320.017—
Stabilized Jordan-RoPEEvaluation context length=8192, eta=.0052026.051.9450.0740.004
Stabilized Jordan-RoPEEvaluation context length=8192, eta=.0102026.051.970.0970.0081
Stabilized Jordan-RoPEEvaluation context length=8192, eta=.0202026.052.1310.1770.0176
Stabilized Jordan-RoPEEvaluation context length=8192, eta=.0502026.052.2750.2190.0476
RoPEEvaluation context length=81922026.052.3220.46—
Direct-sumEvaluation context length=81922026.052.5780.6940.0013