Microsoft's highly efficient small model that punches above its weight class, demonstrating the power of quality synthetic data training.
| Benchmark | Description | Score | Category | Bar |
|---|---|---|---|---|
| MMLU | Massive Multitask Language Understanding | 80.3% | knowledge | |
| HumanEval | Code generation accuracy | 79% | coding | |
| MATH | Mathematical reasoning | 75.6% | math | |
| GSM8K | Grade school math | 80.1% | math | |
| GPQA | Graduate-level science Q&A | 31.3% | reasoning | |
| HellaSwag | Common sense reasoning | 80.1% | reasoning | |
| SWE-bench | Software engineering problems | 16.9% | coding | |
| ARC-C | AI2 Reasoning Challenge | 87% | reasoning |