Hallucination Detection on TruthfulQA AR
87.8F1-ScoreCrossHallu
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CrossHalluBackbone=Aya, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 87.8 | 76.46 | 90.63 | |
| CrossHalluBackbone=Phi4-mini, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 87.06 | 79.34 | 92.2 | |
| CrossHalluBackbone=Phi4-mini, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 86.8 | 70.5 | 88.67 | |
| CrossHalluBackbone=Aya, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 86.8 | 72.76 | 88.38 | |
| CrossHalluBackbone=Aya, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 86.25 | 74.79 | 90.23 | |
| CrossHalluBackbone=Allam, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 84.43 | 65.04 | 83.64 | |
| CrossHalluBackbone=Phi4-mini, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 84.25 | 79.07 | 93.44 | |
| CrossHalluBackbone=Silma, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 83.1 | 79.11 | 86.17 | |
| CrossHalluBackbone=Mistral, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 82.53 | 80.42 | 89.49 | |
| CrossHalluBackbone=Allam, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 81.23 | 45.49 | 72.42 | |
| CrossHalluBackbone=Mistral, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 80.5 | 75.42 | 87.2 | |
| CrossHalluBackbone=Allam, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 78.81 | 50.27 | 72.49 | |
| CrossHalluBackbone=Silma, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 78.62 | 74.25 | 80.63 | |
| CrossHalluBackbone=Silma, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 74.58 | 40.28 | 60.88 | |
| CrossHalluBackbone=Mistral, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 74.05 | 64.24 | 76.72 | |
| CrossHalluBackbone=Qwen2.5, Train dataset=TruthfulQA EN, Transfer setting group=Cross-lingual (CL)2026.07 | 54.02 | 68.52 | 56.41 | |
| CrossHalluBackbone=Qwen2.5, Train dataset=HalluScore, Transfer setting group=Cross-domain (CD)2026.07 | 52.59 | 70.06 | 57.2 | |
| CrossHalluBackbone=Qwen2.5, Train dataset=TruthfulQA AR, Transfer setting group=Same-language baseline (BS)2026.07 | 33.71 | 68.46 | 51.83 |