OPT-175B
Meta · 2022-05-03 · 174.6B parameters
OPT-175B is Meta AI's decoder-only Open Pre-trained Transformer checkpoint, released May 3, 2022 under a non-commercial research license with access granted to eligible researchers. Its 96-layer, 12,288-dimensional architecture supports 2,048-token sequences, and Meta's training chronicle reports that the run completed on 300 billion tokens drawn from a corpus of roughly 180 billion tokens.
Benchmark scores
| Benchmark | Score |
|---|---|
| BoolQ | 79.3 |
| GSM8K | 4.0 |
| HellaSwag (Unspecified Scoring Protocol) | 79.1 |