opt-6.7B: benchmark results
Meta OPT is an English-dominant decoder-only pretrained language model using causal language modeling, a GPT-2 byte-level tokenizer and 2,048-token training sequences. Its training corpus contains 180B tokens; this base release remains separate from OPT-IML instruction-tuned models. Provider: Meta. Released 2022-05-03. Access: Open.
Aggregate Rating 1132 ± 16, rank #10284 of 10916 rated models, from 75 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MMLU-by-task - High School Mathematics | 26.3 | Accuracy (%) | 40 |
| MMLU-by-task - Machine Learning | 29.46 | Accuracy (%) | 35.2 |
| MMLU-by-task - Anatomy | 34.81 | Accuracy (%) | 34.9 |
| MMLU-by-task - College Computer Science | 32 | Accuracy (%) | 32.6 |
| MMLU-by-task - Professional Accounting | 28.01 | Accuracy (%) | 32.6 |
| MMLU-by-task - Virology | 32.53 | Accuracy (%) | 32.1 |
| MMLU-by-task - HellaSwag | 50.68 | Accuracy (%) | 31.9 |
| MMLU-by-task - Public Relations | 33.64 | Accuracy (%) | 30.2 |
| MMLU-by-task - Global Facts | 30 | Accuracy (%) | 30 |
| MMLU-by-task - Marketing | 30.34 | Accuracy (%) | 29.8 |
| MMLU-by-task - Human Aging | 34.53 | Accuracy (%) | 28.2 |
| MMLU-by-task - Prehistory | 28.4 | Accuracy (%) | 26.2 |
Interactive version: theaggregate.ai/model?slug=opt-6-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.