SWE-smith

0.0.9 · active · verified Fri Apr 10

SWE-smith is an open-source Python toolkit designed for generating large-scale software engineering training data. It enables users to turn any GitHub repository into a 'SWE-gym' to create unlimited task instances (e.g., file localization, program repair, SWE-bench) for training Software Engineering (SWE) agents. The current version is 0.0.9, and it appears to be actively developed, with frequent updates and an upcoming NeurIPS 2025 Datasets & Benchmarks Track spotlight. [2, 4, 6]

Warnings

breaking SWE-smith relies heavily on Docker for creating and managing execution environments. Lack of Docker or running on unsupported OS (like Windows/MacOS directly) can lead to unexpected behavior or prevent functionality. [4]
Fix: Ensure Docker is installed and running, and preferably use a Linux-based OS like Ubuntu 22.04.4 LTS for development. [4]
gotcha The primary interaction model often involves running specific modules via `python -m swesmith.module.submodule` for tasks like bug generation, validation, or environment building, rather than direct class instantiation and method calls for core workflows. [1, 10]
Fix: Refer to the official documentation and quickstart guides for the correct command-line usage patterns for different SWE-smith workflows. [1, 10]
gotcha While the core library is Python-focused, SWE-smith is expanding to support other programming languages (Go, JavaScript, Rust, C, C++, C#, Java, PHP). Be aware that full functionality and bug generation strategies might differ or be under active development for non-Python languages. [6, 8]
Fix: Check the latest documentation and GitHub repository for specific language support details and available features for your target language. [6, 8]

Install

pip install swesmith PyPI
git clone git@github.com:SWE-bench/SWE-smith.git cd SWE-smith pip install -e '.[all]' From Source (with all dependencies)

Imports

registry
```
from swesmith.profiles import registry
```
This is a common import for accessing repository profiles when working with SWE-smith datasets. [4]

Quickstart

This quickstart demonstrates how to load a SWE-smith dataset using the `datasets` library and retrieve the `RepoProfile` for a given task instance. It outlines the initial steps for interacting with SWE-smith generated data, typically leading to environment creation and agent training. Note that full execution, particularly `rp.get_container(task)`, requires Docker to be running. [4]

# Example: Loading a SWE-smith dataset and getting a RepoProfile
# Requires 'datasets' to be installed (pip install datasets)
import os
from datasets import load_dataset
from swesmith.profiles import registry

# NOTE: This example requires Docker to be running for environment creation
# and may download a large dataset. Authentication (e.g., Hugging Face token)
# might be needed depending on dataset access.

# Load a small sample of the SWE-smith dataset
try:
    ds = load_dataset("SWE-bench/SWE-smith", split="train", streaming=True)
    print("Dataset loaded successfully. Processing first few tasks...")
    
    count = 0
    for task in ds:
        if count >= 2:  # Process only the first 2 tasks for quickstart
            break
        print(f"\n--- Processing Task {count + 1} ---")
        print(f"Task ID: {task.get('instance_id', 'N/A')}")
        
        # Get the RepoProfile for the task
        rp = registry.get_from_inst(task)
        print(f"Repository Profile for task: {rp.repo_name}")
        
        # Get a pointer to a Docker container with the task initialized (requires Docker)
        # This step will actually attempt to create/get a Docker container
        # Skipping actual container interaction for a simple quickstart printout.
        # container = rp.get_container(task)
        # print(f"Container ID for task: {container.id}")
        print("To get the Docker container, uncomment 'container = rp.get_container(task)'")
        count += 1
except Exception as e:
    print(f"An error occurred during quickstart: {e}")
    print("Please ensure Docker is running and 'datasets' is installed. "
          "If using a private dataset, ensure you are logged in (e.g., huggingface-cli login).")

view raw JSON →