Phi-3.5-mini.
Microsoft Aug 2024 Open weights
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SIQA | 74.7% | #5 of 24 | Lab model card | |
| OpenBookQA | 79.2% | #7 of 24 | Lab model card | |
| Qasper | 41.9% | #1 of 3 | Lab model card | |
| SQuALITY | 24.3% | #1 of 3 | Lab model card | |
| TruthfulQA | 64% | #13 of 30 | Lab model card | |
| Vectara Hallucination Leaderboard (HHEM) | 4.10% | #68 of 148 | Official leaderboard | |
| ARC-Challenge | 84.6% | #25 of 49 | Lab model card | |
| MMLU-CF | 58.3% | #22 of 37 | Official leaderboard | |
| BIG-Bench Hard | 69% | #50 of 73 | Lab model card | |
| PIQA | 81% | #26 of 37 | Lab model card | |
| RULER | 84.1% | #31 of 42 | Lab model card | |
| BoolQ | 78% | #21 of 28 | Lab model card | |
| MMLU | 69% | #102 of 131 | Lab model card | |
| EQ-Bench v2 | 54.7 | #74 of 85 | Official leaderboard | |
| WinoGrande | 68.5% | #41 of 47 | Lab model card | |
| MMMLU | 55.4% | #64 of 71 | Lab model card | |
| HellaSwag | 69.4% | #49 of 53 | Lab model card | |
| MGSM | 47.9% | #53 of 57 | Lab model card | |
| MMLU-Pro | 47.5% | #157 of 167 | Lab model card | |
| Arena-Hard-Auto | 37% | #38 of 39 | Lab model card | |
| RepoQA | 77% | #4 of 4 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GSM8K | 86.2% | #55 of 92 | Lab model card | |
| KOR-Bench | 39% | #21 of 33 | Official leaderboard | |
| MATH | 48.5% | #72 of 99 | Lab model card | |
| ZebraLogic | 6.40% | #43 of 46 | Official leaderboard | |
| GPQA Diamond | 27.2% | #298 of 302 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MBPP | 68.6% | #47 of 71 | Lab model card | |
| BigCodeBench-Hard | 14.8% | #70 of 96 | Official leaderboard | |
| BigCodeBench | 35.6% | #55 of 69 | Official leaderboard | |
| HumanEval | 61.5% | #81 of 101 | Lab model card |
The Self-Reported Index fits one score to the 20 capability results Microsoft reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 108.9 to 114.2. How we compute it.
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
Every result in the longer documents the lab published with it. 23 of them aren't in the launch post.