---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- code
- software-engineering
- swe-bench
- harbor
- evaluation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: preview
    path: preview.csv
---

# Monthly-SWEBench-2026-03

**Monthly-SWEBench-2026-03** is a curated benchmark of 112 real-world software engineering tasks, sourced from GitHub pull requests merged in March 2026. Tasks are in [Harbor](https://github.com/laude-institute/harbor) format and can be run with any Harbor-compatible agent.

[View leaderboard and results →](https://unipat.ai/benchmarks/MonthlySWEBench)

- **112 tasks** — 68 bugfix + 44 non-bugfix
- Tasks span diverse open-source repositories
- Each task includes a runnable environment, test suite, and reference solution

---

## Task Structure

Each task is a self-contained directory:

```
<task_id>/
├── task.toml        # Task metadata and resource limits
├── instruction.md   # Natural-language description given to the agent
├── environment/     # Dockerfile + setup scripts defining the starting state
├── tests/           # Verifier scripts run after the agent finishes
└── solution/        # Reference solution (ground-truth patch)
```

**`task.toml`** specifies the task name, difficulty, category, and execution constraints — agent timeout, verifier timeout, CPU/memory/disk limits for the container.

**`instruction.md`** is what the agent sees. It describes what needs to be implemented or fixed, the expected behavior, and acceptance criteria — written as a natural language engineering spec, not as a hint or walkthrough.

**`environment/`** contains the Dockerfile (and any supporting files) that builds the container the agent works inside. The environment captures the exact repository state before the PR was merged — the agent starts from a broken or incomplete codebase.

**`tests/`** contains the verifier. After the agent finishes, Harbor runs the test script inside the container. The script writes `1.0` to `/logs/verifier/reward.txt` if the task passes, `0.0` if it fails. Tests are derived from the actual CI checks in the original PR.

**`solution/`** contains the reference patch — the actual code change from the merged PR. This is used by the oracle agent and can be used for debugging or analysis.

---

## Run with Harbor

### Install

```bash
pip install harbor
```

### Download

```bash
huggingface-cli download UnipatAI/Monthly-SWEBench-2026-03 \
  --repo-type dataset --local-dir ./monthly-swebench-2026-03

cd monthly-swebench-2026-03
tar -I zstd -xf bugfix.tar.zst
tar -I zstd -xf non_bugfix.tar.zst
```

### Run

```bash
export ANTHROPIC_API_KEY=<your-key>

# Run bugfix tasks
harbor run \
  -p ./monthly-swebench-2026-03/bugfix \
  --agent terminus-2 \
  --model anthropic/claude-opus-4-5 \
  --n-concurrent 4

# Run non-bugfix tasks
harbor run \
  -p ./monthly-swebench-2026-03/non_bugfix \
  --agent terminus-2 \
  --model anthropic/claude-opus-4-5 \
  --n-concurrent 4
```

`terminus-2` is Harbor's built-in multi-turn agent and works with any model supported by LiteLLM. Swap `--model` to evaluate different models, or use `--agent` to try other agents (`claude-code`, `codex`, `openhands`, `aider`, etc.).

### View results

```bash
harbor view
```

---

## Citation

```bibtex
@dataset{monthly_swebench_2026_03,
  title  = {Monthly-SWEBench-2026-03},
  author = {UnipatAI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/UnipatAI/Monthly-SWEBench-2026-03}
}
```