Question Answering on SQuAD (held-out)
50.9Exact Match (EM)GPT-2 small (fine-tuned)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GPT-2 small (fine-tuned)ARACH=true, Logit offset (b)=-0.32026.03 | 50.9 | 60.48 | |
| GPT-2 small (fine-tuned)ARACH=true, Logit offset (b)=-0.42026.03 | 50.73 | 60.32 | |
| GPT-2 small (fine-tuned)ARACH=true, Logit offset (b)=-0.52026.03 | 50.54 | 60.18 | |
| GPT-2 small (fine-tuned)ARACH=false2026.03 | 50.41 | 59.71 |