Kind: Interface
Source: atloria-monorepo/libs/agent-core/src/benchmark/swebench-subset.ts
SWE-bench Verified — Curated Starter Subset
30 hand-picked tasks from the actual SWE-bench Verified dataset (500 tasks). Balanced 10/10/10 across difficulty for clean comparison. Selected for: well-defined test targets (FAIL_TO_PASS), pure Python, and coverage across 9 repos.
Difficulty based on gold patch line count: easy = < 20 lines medium = 20-80 lines hard = > 80 lines
IDs verified against princeton-nlp/SWE-bench_Verified (2026-03-25).
Properties
| Property | Type |
|---|---|
id | string |
repo | string |
difficulty | `'easy' |
description | string |
Was this page helpful?