Commonsense Reasoning on HellaSwag (first 1000 examples)
38Accuracy (HellaSwag 1k)k=64 (DR only)
Evaluation Results
| Method | Links | |
|---|---|---|
| k=64 (DR only)Model size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=1000, Depth Registers=k=642026.04 | 38 | |
| k=64 + sinkModel size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=1000, Depth Registers=k=64, Attention-sink loss=included2026.04 | 36.5 | |
| k=0 (baseline)Model size=300M, Tokens=5B, Precision=FP16, Decoding=greedy, Mode=zero-shot, Examples=10002026.04 | 35.7 |