Open-ended Generative QA on TruthfulQA (ACC, ECE)
94.6Accuracy (ACC)UQ4CT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| UQ4CTBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=1, Correctness proxy=F1-based2024.10 | 94.6 | 6.08 | |
| Verbalized ConsistencyBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=5, Correctness proxy=F1-based2024.10 | 87.49 | 11.35 | |
| Semantic EntropyBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=5, Correctness proxy=F1-based2024.10 | 86.66 | 18.19 | |
| Semantic DensityBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=5, Correctness proxy=F1-based2024.10 | 86.66 | 16.65 | |
| VerbalizedBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=1, Correctness proxy=F1-based2024.10 | 85.63 | 13 | |
| SeqProbBackbone=Llama-3.1-8B-Instruct, Fine-tuning=MixLoRA, Number of samples (N)=1, Correctness proxy=F1-based2024.10 | 84.19 | 9.93 |