Unverified50% confidenceFactExact time

The SWE-bench paper reports that its task repositories average 438,000 lines of code, far exceeding the model's context window.

1

Sources

50%

Confidence

Long-term

Relevance

10/9/2026

First Seen

Sources

Inside the Agent Loop: How Claude Code and OpenAI Codex Actually Work

Related Entities

SWE-benchconcept

Related Claims

UnverifiedThe training data of mainstream LLMs contains a large amount of open-source embedded code, such as tens of thousands of STM32 HAL examples on GitHub.60% similar VerifiedCodex's training data covers billions of lines of publicly available code on GitHub.59% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/991582

API

curl https://kongchang.com/api/v1/knowledge/claims/991582

MCP

get_claim(id=991582)