Atlas

Models

← All models

Code Llama 7B

Meta · 2023-08-24 · 6.7B parameters

Code Llama 7B is the smallest base checkpoint of Meta's Code Llama family, released on August 24, 2023. It continues training Llama 2 7B on 500B tokens of code-heavy data with a fill-in-the-middle objective, so it supports both left-to-right completion and infilling, and a long-context stage at 16,384-token sequences with a RoPE base period of 1,000,000 lets Meta's inference code run it on inputs of up to 100,000 tokens.

Benchmark scores

BenchmarkScore
Capability92.4
GSM8K13.0
HumanEval33.5
HumanEval (pass@10)59.6
HumanEval (pass@100)85.9
MBPP41.4
MBPP (pass@10)66.7
MBPP (pass@100)82.5
MultiPL-E HumanEval (Bash)12.0
MultiPL-E HumanEval (C#)25.3
MultiPL-E HumanEval (C++)28.6
MultiPL-E HumanEval (Java)34.2
MultiPL-E HumanEval (PHP)24.2
MultiPL-E HumanEval (TypeScript)33.3
ToxiGen - toxic generation rate22.6
TruthfulQA generation - % informative86.7
TruthfulQA generation - % truthful38.3
TruthfulQA generation - % truthful and informative26.2
Loading Atlas data…