Codex

anchor

Findings

• IC-1231On APPS, released code generation models span pass@1 from 0.20 (GPT-3 175B) to 6.20 (CodeRL), with value-based and policy-based RL methods outperforming supervised baselines