Unverified50% confidenceFactExact time
Mainstream benchmarks like SWE-Bench Pro heavily rely on issues and commits from public repositories, creating data contamination risk for models trained on public GitHub code.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
DeepSWE Benchmark Reveals the Truth: GPT 5.5 Leads Opus 4.7 by a Wide Margin
bilibili7k的每日搬运5/29/2026
Related Claims
UnverifiedSWE-bench评估模型解决真实GitHub issue的能力，LiveCodeBench使用竞赛编程题目防止数据泄露71% similarUnverified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70% similarUnverifiedSWE-bench（Software Engineering Benchmark）是目前衡量智能体编码能力的核心基准，从真实的GitHub Issue出发要求模型自主定位代码库中的问题并提交修复补丁68% similarUnverifiedGitHub Actions 存在平台绑定问题，国内开发者使用 GitHub 托管 Runner 面临网络访问不稳定的风险68% similarUnverifiedGitHub Copilot生成代码中常见的安全问题包括SQL注入、跨站脚本攻击（XSS）和不安全的随机数生成67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/46950API
curl https://kongchang.com/api/v1/knowledge/claims/46950MCP
get_claim(id=46950)