Open-ended Generation on TruthfulQA open-ended (BLEU/ROUGE)
51.2BLEURAb
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RAbArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 51.2 | 49.7 | 41.6 | 46.8 | |
| RAdArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 50.9 | 54.1 | 46.8 | 54.6 | |
| RAdArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 49.5 | 47.7 | 39.5 | 44.3 | |
| RAdArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 47.7 | 53.9 | 40.9 | 51.9 | |
| Base modelArchitecture=Zephyr-7B2026.01 | 47 | 45.5 | 37.9 | 42.6 | |
| RAbArchitecture=Mistral-7B, Concept Direction=random2026.01 | 42.8 | 41.4 | 37.9 | 42 | |
| RAbArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 41.1 | 41.9 | 31.6 | 40.9 | |
| Base modelArchitecture=Mistral-7B2026.01 | 40.6 | 38.7 | 35.5 | 40.6 | |
| RAdArchitecture=Mistral-7B, Concept Direction=random2026.01 | 40.4 | 39.9 | 38.2 | 40.4 | |
| RAbArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 36.2 | 33.8 | 27.9 | 35 |