DeepSeek-Coder-Base 7B v1.5
DeepSeek · 2024-01-25 · 6.9B parameters
DeepSeek-Coder-Base v1.5 is a 7B code model released in January 2024, continued-pretrained from DeepSeek LLM 7B on code, natural-language and math data to keep coding ability while improving math and general language skills. It has the DeepSeek LLM 7B architecture: 6,910,365,696 parameters, 30 layers with multi-head attention and a 4,096-token context.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC-Challenge | 47.2 |
| BBH (BIG-Bench Hard) | 55.2 |
| Capability | 91.8 |
| HellaSwag (Unspecified Scoring Protocol) | 69.9 |
| HumanEval | 43.2 |
| HumanEval-Mul | 43.2 |
| MBPP | 60.4 |
| MMLU | 49.1 |
| WinoGrande | 63.8 |