Unverified50% confidenceFactExact time
For code benchmarks the contamination risk is especially prominent, because contest solutions and discussion posts on code hosting platforms such as GitHub are often included in training corpora.
1
Sources
50%
Confidence
Long-term
Relevance
10/25/2026
First Seen
Sources
Related Entities
Related Claims
Unverified数据污染问题源于大模型在包含 GitHub、Stack Overflow、LeetCode 等平台公开代码题解的语料上进行预训练80% similarUnverifiedSelf-hosted GitLab is the most sensitive risk surface among the three, because a compromise of the code hosting platform itself means source code and CI credentials are exposed at the same time, and the upgrade pace of self-hosted instances is usually far slower than SaaS.77% similarUnverifiedGitHub Copilot's training data comes from public code repositories on GitHub, which has sparked controversy over code copyright and license compliance.74% similarUnverifiedPublic code (such as GitHub) has been widely used by various companies for training, with diminishing marginal value; real developer interaction traces, private repository corpora, and editing context constitute differentiated advantages.74% similarUnverifiedA 2022 study by Stanford University found that developers using GitHub Copilot to assist with programming had a higher probability of producing insecure code in specific tasks than a control group not using assistive tools.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1095818API
curl https://kongchang.com/api/v1/knowledge/claims/1095818MCP
get_claim(id=1095818)