DeepSeek-Coder-Base 6.7B
DeepSeek · 2023-11-02 · 6.7B parameters
DeepSeek-Coder-Base 6.7B is the dense 6.7B base checkpoint of the original DeepSeek-Coder series, released November 2, 2023. It was trained from scratch on 2 trillion tokens: 1.8 trillion at 4K context followed by 200 billion at 16K, with 87% code and 13% natural language, and uses a fill-in-the-middle objective for repository-level code completion and infilling.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC AI2 | 36.4 |
| BigCodeBench Complete | 41.8 |
| BigCodeBench-Hard Complete | 13.5 |
| Capability | 83.9 |
| Epoch Capabilities Index (ECI) | 87.9 |
| GSM8K | 21.3 |
| MMLU | 36.4 |
| WinoGrande | 57.6 |