Atlas

Models

← All models

Code Llama 13B

Meta · 2023-08-24 · 13B parameters

Code Llama 13B is the 13-billion-parameter base checkpoint of Meta's Code Llama family, released on August 24, 2023. It continues training Llama 2 13B on 500B tokens of code-heavy data with a fill-in-the-middle objective, giving both completion and infilling, and its 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
Capability96.9
GSM8K20.8
HumanEval36.0
HumanEval (pass@10)69.4
HumanEval (pass@100)89.8
MBPP47.0
MBPP (pass@10)71.7
MBPP (pass@100)87.1
MultiPL-E HumanEval (Bash)15.2
MultiPL-E HumanEval (C#)27.3
MultiPL-E HumanEval (C++)39.1
MultiPL-E HumanEval (Java)38.0
MultiPL-E HumanEval (PHP)34.2
MultiPL-E HumanEval (TypeScript)29.6
ToxiGen - toxic generation rate22.4
TruthfulQA generation - % informative89.8
TruthfulQA generation - % truthful43.0
TruthfulQA generation - % truthful and informative33.3
Loading Atlas data…