Unverified50% confidenceFactExact time
2023 年一项学术研究发现在特定提示条件下 Codex 模型约 1% 的输出与训练数据存在逐字匹配
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
Rust官方规范AI代码贡献：LLM使用政策解读
Related Claims
Unverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分66% similarUnverified2024 年的多项基准测试表明，经过严格安全对齐的模型在代码生成任务上的拒绝率比基座模型高出 15-30%，在创意写作任务上的自我审查率高达 40%65% similarUnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions65% similarUnverifiedA study published at the ACL conference in 2024 showed that when a particular viewpoint appears in training data more than three times as frequently as the opposing viewpoint, the model tends to present that viewpoint as 'mainstream consensus'.65% similarUnverifiedA 2024 study indicates that even when AI-generated content accounts for only 10–20% of training data, measurable quality degradation can occur after multiple generations of training.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/718420API
curl https://kongchang.com/api/v1/knowledge/claims/718420MCP
get_claim(id=718420)