Qwen1.5 110B
Alibaba · 2024-04-25 · 111.2B parameters
Qwen1.5-110B is the largest base (pretrained) checkpoint of Alibaba's Qwen1.5 family, released on April 25, 2024 as the first 100B+ Qwen model. It is a dense decoder-only Transformer with grouped-query attention, 111,209,914,368 parameters and a 32,768-token context window; Qwen1.5-110B-Chat is its aligned counterpart.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC-Challenge | 69.6 |
| BBH (BIG-Bench Hard) | 74.8 |
| C-Eval | 89.1 |
| Capability | 113.2 |
| CMMLU | 88.3 |
| GPQA - release-reported, subset unspecified | 35.9 |
| GSM8K | 85.4 |
| HellaSwag (Unspecified Scoring Protocol) | 87.5 |
| MATH | 49.6 |
| MMLU | 80.4 |
| MMLU Pro | 49.4 |
| MultiPL-E (HumanEval, 8-language average) | 52.7 |
| TheoremQA | 34.9 |
| TruthfulQA (Unspecified Metric) | 49.6 |
| WinoGrande | 83.5 |