TriviaQA (closed-book, Wikipedia validation)
General QA · 2017-05-09
Closed-book TriviaQA on the Wikipedia-domain validation subset, as used in the Llama 2 paper: the model answers trivia questions without the evidence documents and is scored by exact match against the answer aliases. The same model scores well above its filtered-development-set result here, so the two are kept apart.
Top models (higher is better)
| Model | Score |
|---|---|
| Llama 3.1 405B | 91.8 |
| Llama 3.1 70B | 89.8 |
| Llama 3 70B | 89.7 |
| Llama 3 8B | 78.5 |
| Llama 3.1 8B | 77.6 |