DeepSWE-style repository coding tasks — delivered data line

Original repository-level coding tasks aligned with the DeepSWE difficulty band: bug repair, feature extension, and algorithm implementation in real open-source repositories, each with a bootable Docker environment and a deterministic grader.

What a task package includes

instruction.md — full model-visible brief with constraints;
environment/ — pinned Docker image with the baseline repository;
private hidden tests (F2P/P2P) and a deterministic grader;
golden patch and solve script, stored separately;
execution trajectories including failures;
SHA-256 manifest, provenance (upstream commits, grader hashes), and a validation file reporting counts honestly.

Measured difficulty

Under one fixed budget — GPT-5.6 Sol (xhigh), 4 valid attempts per task — every delivered DeepSWE-style task was solved at most 2 of 4 attempts, and 3 of 5 tasks were never solved. Languages covered: Rust, Python, Go, TypeScript, Shell. One public view of a real task (rustworkx GNM sampler, 5 requirements levels) is on GitHub.

Weekly capacity

200 tasks per week across the five languages above is our delivery capacity statement, not a committed SLA; volume commitments are scoped per project.

Request a sample

Write to contact@datamining.cloud with your capability target; we provide a redacted sample package for evaluation. Related: RL environment delivery overview · difficulty report