TruthfulQA generation - % truthful and informative
Safety · 2021-09-08
TruthfulQA's generation task asks 817 questions written so that repeating common misconceptions gives false answers. Fine-tuned GPT-judge and GPT-info models rate each free-form answer, and this score is the percentage of answers judged both truthful and informative.
Top models (higher is better)
| Model | Score |
|---|---|
| Llama 2 70B Chat | 64.1 |
| Llama 2 13B Chat | 62.2 |
| Llama 2 7B Chat | 57.0 |
| Llama 2 70B | 50.2 |
| Code Llama Instruct 34B | 47.4 |
| Llama 2 34B | 43.5 |
| Llama 2 13B | 41.9 |
| Falcon 40B | 40.4 |
| Code Llama Instruct 13B | 36.8 |
| MPT-30B | 35.3 |
| Code Llama 34B | 34.6 |
| Llama 2 7B | 33.3 |
| Code Llama 13B | 33.3 |
| Code Llama Instruct 7B | 31.5 |
| MPT-7B | 29.1 |
| Code Llama 7B | 26.2 |
| Falcon 7B | 25.9 |