distilgpt2 is the lightweight sibling of GPT-2, trained to mimic its behavior at roughly half the size. It generates fluent short text quickly but loses coherence on longer passages and struggles with complex reasoning. Think of it as a fast sketch artist — useful for prototyping and experimentation, but not for polished output.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| MuSR | 11.2 | accuracy | 4mo ago |
| MATH | 0.6 | accuracy | 4mo ago |
| GPQA Diamond | 1.2 | accuracy | 4mo ago |
| MMLU-Pro | 2.1 | accuracy | 4mo ago |
| IFEval | 6.1 | accuracy | 4mo ago |
| BBH | 2.8 | accuracy | 4mo ago |