Unverified50% confidenceFactExact time
2021年Carlini等人的研究首次系统性证明GPT-2能以较高概率逐字复现训练数据中的个人信息、代码片段及版权文本，并将其定义为训练数据提取攻击
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified研究人员已证明GPT-2可以在特定提示下逐字输出训练集中的真实邮件地址、电话号码甚至个人姓名73% similarUnverified谷歌DeepMind的研究发现GPT类模型在特定提示条件下可精确复现训练集中约1%的文本片段71% similarUnverifiedAccording to official data, GPT-Live 1 shows significant improvements over AVM on benchmarks such as expert-level scientific reasoning and agentic web search.71% similarVerifiedThe GPT series of models are trained on trillion-scale text data through a self-supervised task of predicting the next token.69% similarUnverifiedMultiple academic studies have shown that a considerable proportion of the high scores achieved by mainstream models such as GPT-4 and LLaMA on leaderboards including MMLU, HumanEval, and GSM8K stem from overlap between training set and test set content.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/533822API
curl https://kongchang.com/api/v1/knowledge/claims/533822MCP
get_claim(id=533822)