DeepSeek-V2-Lite-Chat
DeepSeek · 2024-05-16 · 15.7B parameters
DeepSeek-V2-Lite-Chat is the supervised-fine-tuned chat version of DeepSeek-V2-Lite, released in May 2024. It keeps the base model's 15,706,484,224-parameter MLA/DeepSeekMoE architecture (about 2.4B activated per token by DeepSeek's rounded figure) and documented 32K context.
Benchmark scores
| Benchmark | Score |
|---|---|
| AGIEval | 42.8 |
| AlignBench (v1.0 or version unstated) | 6.0 |
| AlpacaEval 2.0 (length-controlled win rate) | 16.9 |
| Arena-Hard-Auto v0.1 | 11.4 |
| BBH (BIG-Bench Hard) | 48.1 |
| C-Eval | 60.1 |
| Capability | 96.9 |
| CLUEWSC | 80.0 |
| CMATH | 71.7 |
| CMMLU | 62.5 |
| GSM8K | 72.0 |
| HumanEval | 57.3 |
| MATH | 27.9 |
| MBPP | 45.8 |
| MMLU | 55.7 |
| MT-Bench | 7.4 |
| Natural Questions (open-domain, closed-book) | 35.5 |
| TriviaQA | 65.2 |