CommonSenseQA
General QA · 2018-11-02
CommonSenseQA (version 1) is a five-way multiple-choice question-answering benchmark built from ConceptNet relations, where answering requires everyday commonsense knowledge rather than a supporting passage. Test labels are hidden, so results are usually accuracy on the 1,221-question development set. It is a different dataset from CommonSenseQA 2.0.
Top models (higher is better)
| Model | Score |
|---|---|
| Llama 3.1 405B | 85.8 |
| Llama 3.1 70B | 84.1 |
| Llama 3 70B | 83.8 |
| Mixtral 8x22B | 82.4 |
| Llama 3.1 8B | 75.0 |
| Gemma 7B | 74.4 |
| Llama 2 34B | 74.3 |
| LLaMA 65B | 74.0 |
| Llama 3 8B | 72.6 |
| LLaMA 33B | 72.5 |
| Falcon 40B | 70.4 |
| LLaMA-13B | 62.0 |
| MPT-30B | 58.2 |
| LLaMA 7B | 33.6 |
| MPT-7B | 21.3 |
| Falcon 7B | 20.8 |