LLaMA 65B
Meta · 2023-02-24 · 65.3B parameters
LLaMA 65B is Meta's original pretrained base model and the largest checkpoint in the first LLaMA family, publicly announced on February 24, 2023. Meta reported it as a 65.2-billion-parameter model trained on a rounded 1.4 trillion tokens and distributed the weights to approved research applicants under a non-commercial bespoke license.
Benchmark scores
| Benchmark | Score |
|---|---|
| Artificial Analysis Intelligence Index | 2.1 |
| Capability | 106.0 |
| Epoch Capabilities Index (ECI) | 109 |
| HellaSwag (Unspecified Scoring Protocol) | 84.2 |
| LAMBADA | 77.7 |
| MMLU | 63.4 |
| OpenBookQA | 60.2 |
| PIQA | 82.8 |
| WinoGrande | 77.0 |