Software engineering training data

Build in real codebases.Ship changes that hold.

Train on production repositories and full toolchains. Implement features, fix defects, and complete multi-PR changes with deterministic verification.

Try 5 Samples
3,000tasks available
Fulltoolchains
6task families
Real repository
  • src/
  • tests/
  • toolchain/
Tested change
Runnable environments. Verified rewards.

The DeepSWE catalog

Choose your engineering work.

Build and connect

Deliver new capabilities

3task families
  • Feature implementation
  • Multi-PR changes
  • Integrations

Add capabilities to real codebases. Land large changes in reviewable steps and make systems work together.

Diagnose and improve

Make existing code better

3task families
  • Bug fixes
  • Incident response
  • Refactoring

Find the root cause, repair the defect, and improve working code while preserving the behavior users rely on.

Choose your task families and the volume you need across 3,000 available DeepSWE tasks.

Inside a DeepSWE task: pytest

A passing claim isn't a passing test.

Fix a real pytest bug that drops chained exception details during parallel test runs. Restore the full failure output.

Parallel testsDropped exception chains
Fix + verifyRestore the failure output
Clean test runHeld-out repository tests
Accurate updatesManager checks actual work

Completion must match the changed files and the test results.Task workflow

Read the task and verification summary

Repair a pytest bug that drops chained exception details during parallel test runs. The agent must restore the full failure output and verify the change against the repository’s tests.

Held-out repository tests check the fix, and manager checks compare the agent’s completion updates with its actual work. The reference solution must score 1.0 and an empty diff must score 0.0. Both controls run before shipping and are included with the task.

Request the complete sample package ↗

Included with your tasks

Everything needed to run the work.

Inspect the complete task package in your samples.

Task specification
Runnable environment
Reward verifier
Reference solution

Deterministic verification

Real changes. Verified rewards.

A reference solution must score 1.0. An empty diff must score 0.0. Both controls run before shipping and are included with the task.

Fulltoolchain runs
Finalrepository checked
Explore the full catalog ↗
1.0reference reward
Reference solution1.0
Empty diff0.0

Your DeepSWE mix. Your volume.

Choose your engineering task families and the task volume you need.

1Review samples2Choose mix + volume3Get started now