SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Skywork-SWE-32B - Kunlun Wanwei's Open-Source Autonomous Code Agent Base Model

Author: AI小集 · Original source: AI工具集

What is Skywork-SWE-32B

Skywork-SWE-32B is Kunlun Wanwei's open-source 32B-scale software engineering (SWE) autonomous code agent base model. The model focuses on software engineering tasks, particularly repository-level code repair capabilities, and performs excellently in complex scenarios involving multi-turn interaction and long-text processing. By constructing over 10,000 verifiable GitHub repository task instances, it has built the largest verifiable GitHub repository-level code repair dataset to date, achieving a 38.0% pass@1 accuracy on the SWE-bench Verified benchmark, setting a new best result for models of the same parameter scale. After introducing test-time scaling techniques, the accuracy is further improved to 47.0%, significantly surpassing existing open-source models under 32B and approaching or even exceeding the performance of some closed-source models.

Skywork-SWE-32B

Main Features of Skywork-SWE-32B

  • Repository-Level Code Repair: It can locate code issues (such as bugs) in GitHub repositories, generate repair code, verify the repair effect, and complete the full closed loop from problem understanding to resolution.
  • Multi-Turn Interaction Capability: It supports more than 50 turns of interaction, simulating the multiple debugging and repair processes in real development scenarios to gradually solve problems.
  • Long-Text Processing: It can process long texts exceeding 32k tokens, meeting the processing needs of complex code files and multi-file dependencies.
  • Automated Verification: By constructing dedicated runtime environments and unit test verification mechanisms, it ensures that the generated repair code is effective in the actual runtime environment.

Technical Principles of Skywork-SWE-32B

  • Large-Scale Dataset Construction
    • Automated Data Collection and Verification: Through a three-stage automated process (data collection and pre-screening, execution-based verification, and agent trajectory generation), a dataset containing 10,169 real Python task instances was constructed, covering 2,531 different GitHub repositories.
    • Runtime Environment Support: Each task instance is equipped with a dedicated Docker runtime environment image, supporting automated unit test verification to ensure that the generated repair code is effective in the actual runtime environment.
    • High-Quality Training Trajectories: Through the multi-turn interaction trajectories of the agent during task solving, high-quality training samples are generated for model fine-tuning.
  • Model Training and Optimization
    • Based on the OpenHands Framework: It adopts the OpenHands code agent framework, supporting multi-turn interaction and long-text processing, and can simulate the code repair process in real development scenarios.
    • Data Scaling Law: Through systematic verification, it was found that as the scale of training data increases, model performance continues to improve, verifying the applicability of the data scaling law in software engineering tasks.
    • Test-Time Scaling (TTS): In the inference stage, by increasing the number of independent rollouts (such as N=8), the model's performance is further improved, fully utilizing the model's reasoning capability.

Project Address of Skywork-SWE-32B

  • HuggingFace Model Repository：https://huggingface.co/Skywork/Skywork-SWE-32B
  • Technical Paper：https://huggingface.co/Skywork/Skywork-SWE-32B/resolve/main/assets/Report.pdf

Application Scenarios of Skywork-SWE-32B

  • Optimizing Code Quality: The model can analyze potential issues in code, propose optimization suggestions, and help developers improve code quality and maintainability.
  • Unit Test Automation: By building a dedicated runtime environment and a unit test validation mechanism, Skywork-SWE-32B can automatically execute test cases to verify whether the generated fix code is effective.
  • Teaching Assistance: In software engineering and programming courses, Skywork-SWE-32B can serve as a teaching tool to help students understand the process of solving code problems and improve their programming skills.
  • Research Support: It provides researchers with a powerful experimental platform for exploring the application of large language models in software engineering tasks and validating theories such as data scaling laws.
  • Internal Development Tools: Enterprises can integrate Skywork-SWE-32B into internal development tools to automatically handle code problems, reduce manual intervention, and improve development efficiency and code quality.