Character Count Evaluation on Qwen3-8B 200 Prompts (test)
98Accuracy9-row repair
Evaluation results
| Method | Links | |
|---|---|---|
| 9-row repairIntervention type=9-row repair, Argmax restriction=digit-restricted, Number of seeds=3, Training protocol=9-row repair2026.05 | 98 | |
| Probe-roundIntervention type=Probe-round, Argmax restriction=digit-restricted, Number of seeds=32026.05 | 96.8 | |
| Fullvocab repairIntervention type=Fullvocab repair, Argmax restriction=full-vocabulary, Number of seeds=3, Training protocol=margin-hinge training2026.05 | 57.7 | |
| BaselineIntervention type=None, Argmax restriction=digit-restricted, Number of seeds=32026.05 | 49.3 |