Hallucination Detection on RAGTruth Span-level, leakage-clean protocol
0.702AUCGASP+base (trained)
Evaluation Results
| Method | Links | |
|---|---|---|
| GASP+base (trained)Scorer=SmolLM2-1.7B, Protocol=leakage-clean2026.07 | 0.702 | |
| GASP-threshold (default)Scorer=SmolLM2-1.7B, Protocol=leakage-clean2026.07 | 0.681 | |
| GASP+base (trained)Scorer=Qwen2.5-1.5B, Protocol=leakage-clean2026.07 | 0.678 | |
| GASP+base (trained)Scorer=Qwen2.5-0.5B, Protocol=leakage-clean2026.07 | 0.676 | |
| GASP-threshold (default)Scorer=Qwen2.5-1.5B, Protocol=leakage-clean2026.07 | 0.673 | |
| GASP-threshold (default)Scorer=Qwen2.5-0.5B, Protocol=leakage-clean2026.07 | 0.672 | |
| GASP-trainedScorer=SmolLM2-1.7B, Protocol=leakage-clean2026.07 | 0.657 | |
| GASP-trainedScorer=Qwen2.5-1.5B, Protocol=leakage-clean2026.07 | 0.645 | |
| GASP-trainedScorer=Qwen2.5-0.5B, Protocol=leakage-clean2026.07 | 0.635 | |
| PerplexityScorer=SmolLM2-1.7B, Protocol=leakage-clean2026.07 | 0.615 | |
| PerplexityScorer=Qwen2.5-1.5B, Protocol=leakage-clean2026.07 | 0.565 | |
| LengthScorer=Qwen2.5-1.5B, Protocol=leakage-clean2026.07 | 0.548 | |
| LengthScorer=Qwen2.5-0.5B, Protocol=leakage-clean2026.07 | 0.547 | |
| PerplexityScorer=Qwen2.5-0.5B, Protocol=leakage-clean2026.07 | 0.536 | |
| LengthScorer=SmolLM2-1.7B, Protocol=leakage-clean2026.07 | 0.523 |