mpt-30B: benchmark results
MosaicML's MPT-30B, a decoder-style transformer pretrained from scratch on 1 trillion tokens of English text and code, with an 8K context extendable via ALiBi, sized to run on a single GPU. Provider: Databricks. Released 2023-06-22.
Aggregate Rating 1303 ± 11, rank #1476 of 1656 rated models, from 120 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM Classic - Entity Data Imputation | 85.52 | Exact Match (%) | 100 |
| HELM Classic - LSAT | 24.78 | Exact Match (%) | 97.1 |
| HELM Classic - WikiFact | 36.93 | Exact Match (%) | 92.4 |
| MMLU-by-task - Machine Learning | 45.54 | Accuracy (%) | 89.5 |
| HELM Classic - Entity Matching | 89.74 | Exact Match (%) | 89.4 |
| HELM Classic - RAFT | 72.27 | Exact Match (%) | 89.4 |
| HELM Classic - IMDB | 95.9 | Exact Match (%) | 87.9 |
| HELM Classic - NarrativeQA | 73.15 | F1 (%) | 86.2 |
| HELM Classic - bAbI | 54.97 | Exact Match (%) | 81.9 |
| HELM Classic - NaturalQuestions Open Book | 67.29 | F1 (%) | 81.5 |
| MMLU-by-task - College Mathematics | 36 | Accuracy (%) | 80.2 |
| HELM Classic - MATH | 17.81 | Equivalent (%) | 79.4 |
Interactive version: theaggregate.ai/model?slug=mpt-30b · How It Works · Data refreshed daily, snapshot 2026-10-20.