Unverified50% confidenceFactExact time
GPT、Claude、Llama等主流模型的训练语料以英语为主，占比可高达50%-90%
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
UnverifiedEnglish accounts for more than 90% of GPT-4's training data.79% 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%71% similarUnverifiedThe training corpora of mainstream large language models such as the GPT series and Claude are primarily crawled from internet content dating from the 2000s onward.68% similarVerifiedLarge language models such as GPT and BERT are based on the self-supervised learning paradigm68% similarUnverifiedGPT-4、Claude等通用大模型在预训练阶段学习的是互联网上的海量公开文本，倾向于给出通用的回答66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/532878API
curl https://kongchang.com/api/v1/knowledge/claims/532878MCP
get_claim(id=532878)