Qwen2.5 14B
Alibaba · 2024-09-19 · 14.8B parameters
Qwen2.5-14B is a dense base (pretrained) checkpoint of Alibaba's Qwen2.5 series, released on September 19, 2024. It has 14,770,033,664 parameters with grouped-query attention, and Qwen documents a 131,072-token context length; Qwen2.5-14B-Instruct is its instruction-tuned counterpart.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC-Challenge | 67.3 |
| BBH (BIG-Bench Hard) | 78.2 |
| Capability | 115.2 |
| GPQA - release-reported, subset unspecified | 32.8 |
| GSM8K | 90.2 |
| HellaSwag (Unspecified Scoring Protocol) | 84.3 |
| HumanEval | 56.7 |
| HumanEval+ | 51.2 |
| MATH | 55.6 |
| MBPP | 76.7 |
| MBPP+ | 63.2 |
| MMLU | 79.7 |
| MMLU Pro | 51.2 |
| MMLU STEM | 76.4 |
| MMLU-Redux | 76.6 |
| MultiPL-E (HumanEval, 8-language average) | 53.5 |
| TheoremQA | 43.0 |
| TruthfulQA (Unspecified Metric) | 58.4 |
| WinoGrande | 81.0 |