List your company

Character-Level Language Modeling on text8 (held-out 1M tokens)

2.3BPCSHARP
2.2742.44952.6252.8005May 30, 2026

Evaluation results

MethodLinks
SHARPT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=α^L, Params=9.7M (5M active), Wall-time / 1k Tokens (sec)=9.46¹2026.052.3
SHARPT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=α^L, Params=9.7M (5M active), Wall-time / 1k Tokens (sec)=9.46¹2026.052.32
GRUT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=7.3M, Wall-time / 1k Tokens (sec)=14.292026.052.38
Transformer 5MT (BPTT window or input length)=1024, Time Complexity=O(T^2), Effective Context=T, Params=5.0M, Wall-time / 1k Tokens (sec)=116.402026.052.38
GRUT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=7.3M, Wall-time / 1k Tokens (sec)=14.292026.052.41
Transformer 5MT (BPTT window or input length)=1024, Time Complexity=O(T^2), Effective Context=T, Params=5.0M, Wall-time / 1k Tokens (sec)=116.402026.052.41
Transformer 10MT (BPTT window or input length)=1024, Time Complexity=O(T^2), Effective Context=T, Params=9.7M, Wall-time / 1k Tokens (sec)=266.532026.052.41
Transformer 10MT (BPTT window or input length)=1024, Time Complexity=O(T^2), Effective Context=T, Params=9.7M, Wall-time / 1k Tokens (sec)=266.532026.052.43
RNNT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=2.4M, Wall-time / 1k Tokens (sec)=4.412026.052.45
RNNT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=2.4M, Wall-time / 1k Tokens (sec)=4.412026.052.47
LSTMT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=9.7M, Wall-time / 1k Tokens (sec)=20.502026.052.5
LSTMT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=9.7M, Wall-time / 1k Tokens (sec)=20.502026.052.61
Clockwork RNNT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=4.3M, Wall-time / 1k Tokens (sec)=4.912026.052.92
Clockwork RNNT (BPTT window or input length)=4, Time Complexity=O(T), Effective Context=T, Params=4.3M, Wall-time / 1k Tokens (sec)=4.912026.052.95