mpt-30B Instruct: benchmark results
MosaicML's MPT-30B-Instruct, an instruction fine-tune of the MPT-30B decoder-style transformer pretrained on 1 trillion tokens of English text and code, with an 8K ALiBi-extended context. Provider: Databricks. Released 2023-06-22.
Aggregate Rating 1304 ± 12, rank #1470 of 1656 rated models, from 108 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM Classic - NaturalQuestions Open Book | 69.72 | F1 (%) | 93.8 |
| MMLU-by-task - High School Mathematics | 32.96 | Accuracy (%) | 91.8 |
| MMLU-by-task - HellaSwag | 65.25 | Accuracy (%) | 89.8 |
| MMLU-by-task - College Physics | 35.29 | Accuracy (%) | 88.7 |
| HELM Classic - Entity Matching | 89.54 | Exact Match (%) | 87.9 |
| HELM Classic - NarrativeQA | 73.29 | F1 (%) | 87.7 |
| HELM Classic - GSM8K | 34.4 | Exact Match (%) | 86.8 |
| HELM Classic - Synthetic Reasoning Natural | 32.98 | F1 (%) | 86.8 |
| HELM Classic - BoolQ | 85 | Exact Match (%) | 86.4 |
| HELM Classic - MATH Chain-of-Thought | 12.54 | Equivalent (%) | 80.9 |
| HELM Classic - Synthetic Reasoning Abstract | 34.95 | Exact Match (%) | 80.9 |
| MMLU-by-task - College Computer Science | 47 | Accuracy (%) | 80.1 |
Interactive version: theaggregate.ai/model?slug=mpt-30b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.