SWE-rebench-07-2026

SWE-rebench-07-2026 is a curated Harbor dataset containing 111 hard software engineering tasks from the July 2026 SWE-rebench leaderboard snapshot. It is a five-language subset prepared for reproducible agent evaluation in Harbor.

Dataset overview

Property	Value
Tasks	111
Languages	Go (21), Java (21), Python (20), Rust (25), TypeScript (24)
Snapshot	July 2026
Task type	Repository-level issue resolution
Evaluation	Repository-specific tests with SWE-rebench log parsers
Container images	docker.io/swerebenchv2/*

Each task provides an issue-style instruction, a pinned repository environment, an Oracle solution, and a verifier. The task configurations retain the selected language, image_name, and install_config.log_parser fields while omitting intermediate dataset-construction metadata.

Run with Harbor

Run Oracle Setup across the private dataset with ten concurrent workers:

uvx harbor run   -d ibragim-badertdinov/swe-rebench-07-2026@2026-07   -a oracle   --n-concurrent 10

Run a single task:

uvx harbor run   -t ibragim-badertdinov/apache__dubbo-go-3357   -a oracle

Provenance

This Harbor package is derived from the Nebius SWE-rebench leaderboard dataset.

• SWE-rebench V2 dataset
• SWE-rebench leaderboard
• SWE-rebench paper

The source dataset declares the CC BY 4.0 license. Source repositories and their contents remain subject to their respective licenses.

Author

Ibragim Badertdinov — badertdinov.ibragim@gmail.com