Llama 2 34B
Meta
Llama 2 34B is Meta's 34-billion-parameter dense language model from the Llama 2 family, reported in the Llama 2 paper but not publicly released as downloadable weights. Meta described 34B base and chat evaluation results while noting the 34B variants were withheld pending additional red-teaming. This row represents the paper-reported checkpoint used in first-party benchmark tables, not an open-weight release.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC AI2 | 54.5 |
| BBH (BIG-Bench Hard) | 44.1 |
| BoolQ | 83.7 |
| Epoch Capabilities Index (ECI) | 104 |
| GSM8K | 42.2 |
| MMLU | 62.6 |
| OpenBookQA | 58.2 |
| PIQA | 81.9 |
| WinoGrande | 76.7 |