HMM symbol prediction on HMM 3-state v1 (val)
0.005Excess over BayesTwo-layer causal transformer
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Two-layer causal transformerWidth=128, Attention heads=4, MLP width=256, Parameters=275k, Sequence length=64, Training sequences=60002026.05 | 0.005 | 0.402 | 48.4 | 0.107 | 0.03 | 0.004 |