Microsoft·released 2023-12-13✓ 2 sources
Phi-2 benchmark scores: 8 benchmarks tracked. 0% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 8 benchmarks·best result 67.87 on ARC AI2 (self-reported)·0 of 8 independently reproduced
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| ARC AI2 | 67.87accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| OpenBookQA | 64.8accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| BBH | 45.87accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| TriviaQA | 45.2accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| MMLU | 44.53accuracy (%) | unverified· optimizedT1 | 2023-12-12Stanford CRFM Leaderboard |
| HellaSwag | 38.13accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| ANLI | 13.75accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| Winogrande | 9.4accuracy (%) | self-reported· optimizedT1 | 2023-12-12Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
Claims drawn from cited facts, not live model generation.
Developed by Microsoft and released in 2023, this model sits at the compact end of the scale with 2.78 billion parameters. That compact size buys efficiency and lower deployment cost, but it also means less headroom for hard tasks than a large or frontier-scale model would offer.
4 cited facts
The model is scored on 8 tracked benchmarks, and it holds no current top score. Its record is self-reported, meaning the scores are vendor-claimed and not yet independently confirmed, so they should be read as claims rather than verified results.
3 cited facts
On average it trails the best reported score by 39.35 points, a wide gap under the disclosed harness that places it well behind the leaders rather than at the frontier or merely competitive. It holds the top score on 0 benchmarks, meaning the results show no measured category leadership.
2 cited facts
2 facts cross-checked across data sources: 1 corroborated, 1 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: epoch_benchmarks, huggingface_models
Phi-2 is an AI model developed by Microsoft, released 2023-12-13. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Phi-2 has recorded scores on 8 benchmarks, each shown with its evidence status.
Phi-2 has recorded scores on 8 benchmarks — OpenBookQA, Winogrande, TriviaQA, MMLU, ARC AI2, BBH, and 2 more. The full table above shows each score with its evidence status.
0 of 8 recorded scores (0%) are independently reproduced rather than self-reported by the lab.
Phi-2 has 8 tracked claims: 7 self-reported, 1 unverified.