Hallucination Detection on SQuAD (ROC, RAC, F1@B, RA@50)
83.9ROCHaluNet
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HaluNetBackbone=Llama3-8B, CR=1, Sample Latency=97.663, Inference Latency=0.1242025.12 | 83.9 | 94.1 | 53.2 | 96.5 | |
| HaluNetBackbone=Qwen3-14B, CR=1, Sample Latency=128.006, Inference Latency=0.1192025.12 | 78.1 | 96.1 | 39.3 | 97.1 | |
| SelfCheckGPTBackbone=Llama3-8B, CR=1, Sample Latency=760.002, Inference Latency=87.1612025.12 | 77.2 | 87.8 | 46.6 | 94.4 | |
| SEUBackbone=Llama3-8B, CR=1, Sample Latency=731.105, Inference Latency=2.0682025.12 | 76.3 | 7.9 | 45.7 | 6.1 | |
| EmbVarBackbone=Llama3-8B, CR=1, Sample Latency=727.163, Inference Latency=0.1092025.12 | 75.7 | 88.7 | 46.8 | 93.9 | |
| SEUBackbone=Qwen3-14B, CR=1, Sample Latency=1302.565, Inference Latency=1.9482025.12 | 74.8 | 2.4 | 35.6 | 2.3 | |
| P(True)Backbone=Qwen3-14B, CR=1, Sample Latency=118.871, Inference Latency=53.5932025.12 | 74.8 | 97 | 29.2 | 98.5 | |
| LogisticBackbone=Llama3-8B, CR=1, Sample Latency=98.849, Inference Latency=0.0282025.12 | 73.5 | 88.5 | 48.8 | 90.4 | |
| PEBackbone=Llama3-8B, CR=1, Sample Latency=105.435, Inference Latency=0.0012025.12 | 73.2 | 89.8 | 42.7 | 93.4 | |
| EmbVarBackbone=Qwen3-14B, CR=1, Sample Latency=1289.101, Inference Latency=0.1262025.12 | 72.6 | 95.7 | 30.8 | 97.6 | |
| PEBackbone=Qwen3-14B, CR=1, Sample Latency=127.035, Inference Latency=0.0012025.12 | 70.5 | 93.8 | 27.4 | 96.1 | |
| LogisticBackbone=Qwen3-14B, CR=1, Sample Latency=121.403, Inference Latency=0.0422025.12 | 69.9 | 92.4 | 32.2 | 94.3 | |
| P(True)Backbone=Llama3-8B, CR=1, Sample Latency=81.557, Inference Latency=30.4132025.12 | 69.2 | 87.8 | 39.8 | 90.3 | |
| SEBackbone=Qwen3-14B, CR=1, Sample Latency=1253.741, Inference Latency=24.1002025.12 | 66.5 | 95.6 | 27.8 | 96.8 | |
| SelfCheckGPTBackbone=Qwen3-14B, CR=1, Sample Latency=1267.408, Inference Latency=74.0152025.12 | 65.2 | 95.6 | 37.1 | 96.6 | |
| T-NLLBackbone=Llama3-8B, CR=1, Sample Latency=80.780, Inference Latency=0.0012025.12 | 62.8 | 85.4 | 34.7 | 86.8 | |
| SEBackbone=Llama3-8B, CR=1, Sample Latency=743.258, Inference Latency=29.9272025.12 | 61.9 | 84.9 | 34.6 | 85.9 | |
| T-NLLBackbone=Qwen3-14B, CR=1, Sample Latency=131.44, Inference Latency=0.0012025.12 | 57.7 | 94.7 | 13 | 95.6 |