Code Llama 70B
Meta · 2024-01-29 · 69B parameters
Code Llama 70B is the largest base checkpoint of Meta's Code Llama family, released on January 29, 2024. It continues training Llama 2 70B on 1T tokens of code-heavy data with a fill-in-the-middle objective, and a long-context stage with a RoPE base period of 1,000,000 lets it run on inputs of up to 100,000 tokens, unlike the 4,096-token Python and Instruct 70B variants.
Benchmark scores
| Benchmark | Score |
|---|---|
| Capability | 107.0 |
| HumanEval | 53.0 |
| HumanEval (pass@10) | 84.6 |
| HumanEval (pass@100) | 96.2 |
| MBPP | 62.4 |
| MBPP (pass@10) | 81.1 |
| MBPP (pass@100) | 91.9 |
| MultiPL-E HumanEval (Bash) | 29.1 |
| MultiPL-E HumanEval (C#) | 38.0 |
| MultiPL-E HumanEval (C++) | 52.8 |
| MultiPL-E HumanEval (Java) | 51.9 |
| MultiPL-E HumanEval (PHP) | 50.9 |
| MultiPL-E HumanEval (TypeScript) | 49.1 |