Code Llama Python 70B
Meta · 2024-01-29 · 69B parameters
Code Llama - Python 70B is the Python-specialized 70-billion-parameter checkpoint of Meta's Code Llama family, released on January 29, 2024. It continues the 1T-token Code Llama 70B code training with 100B tokens of Python-heavy data but, unlike the base 70B, skips long-context fine-tuning, so it keeps a 4,096-token context and the standard RoPE base period of 10,000.
Benchmark scores
| Benchmark | Score |
|---|---|
| Capability | 107.9 |
| HumanEval | 57.3 |
| HumanEval (pass@10) | 89.3 |
| HumanEval (pass@100) | 98.4 |
| MBPP | 65.6 |
| MBPP (pass@10) | 81.5 |
| MBPP (pass@100) | 91.9 |
| MultiPL-E HumanEval (Bash) | 25.3 |
| MultiPL-E HumanEval (C#) | 34.8 |
| MultiPL-E HumanEval (C++) | 54.7 |
| MultiPL-E HumanEval (Java) | 57.6 |
| MultiPL-E HumanEval (PHP) | 53.4 |
| MultiPL-E HumanEval (TypeScript) | 44.0 |