Unverified60% confidenceBenchmarkExact time
Codex was fine-tuned on top of GPT-3 using approximately 154 billion tokens of publicly available code repository data covering 54 programming languages, achieving a pass@1 accuracy of 28.8% on the HumanEval benchmark, compared to 0% for GPT-3 at the same time.
2
Sources
60%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
Related Claims
UnverifiedCodex是GPT-3的代码微调版本，参数量约12B78% similarUnverifiedGPT-3 的预训练语料以自然语言为主，代码仅占约 1%，而 Codex 的微调数据集中代码占比超过 90%75% similarUnverifiedGPT 6.1 Sol completed the short film project in Codex using high reasoning effort, taking about 12 minutes 27 seconds from request to delivery, with a cumulative 2.755 million tokens, of which 29,884 were output and about 2.634 million were cache reads.73% similarVerifiedGPT-4 and Claude 3 achieve pass@1 rates exceeding 85% on the HumanEval code benchmark.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/492591API
curl https://kongchang.com/api/v1/knowledge/claims/492591MCP
get_claim(id=492591)