README.md

metadata

title: SWE-bench Lite — Eval Explorer
emoji: 🔍
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false

SWE-bench Lite — Eval Explorer

A structural embedding browser over 300 SWE-bench Lite instances. Points are positioned by UMAP applied to AST edit-operation distances — proximity reflects procedural similarity, not semantic content.

What you can do

• Color by fix type, repo, pass/fail, structural coverage, task quadrant, or number of models that solved it
• Filter by fix type, repo, or model outcome (solved by ≥1, solved by all, unsolved)
• Click any point to inspect: per-model outcomes (GPT-4, Claude 3.5, GPT-4o, Qwen2.5), fix summary, structural coverage score, behavioral metrics (steps, edit retry rate)
• Bar chart shows live pass rate by fix type for the current filter

Data

Companion dataset: midah/procedural-info-theory