SWE-Next-14B/README.md

---
license: mit
language:
- en
library_name: transformers
pipeline_tag: text-generation
base_model:
- Qwen/Qwen2.5-Coder-14B-Instruct
datasets:
- TIGER-Lab/SWE-Next-SFT-Trajectories
- TIGER-Lab/SWE-Next
---

<div align="center">
  <h1>SWE-Next: Scalable Real-World Software Engineering Tasks for Agents</h1>
</div>

<div align="center">
  <a href="https://arxiv.org/abs/2603.20691"><img alt="Paper" src="https://img.shields.io/badge/Paper-arXiv-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white"></a>
  <a href="https://tiger-ai-lab.github.io/SWE-Next/"><img alt="Project Page" src="https://img.shields.io/badge/Project%20Page-Website-4285F4?style=for-the-badge&logo=googlechrome&logoColor=white"></a>
  <a href="https://github.com/TIGER-AI-Lab/SWE-Next"><img alt="Code" src="https://img.shields.io/badge/Code-GitHub-181717?style=for-the-badge&logo=github&logoColor=white"></a>
  <a href="https://huggingface.co/datasets/TIGER-Lab/SWE-Next-SFT-Trajectories"><img alt="SFT Trajs" src="https://img.shields.io/badge/SFT%20Trajs-HuggingFace-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000"></a>
  <a href="https://huggingface.co/datasets/TIGER-Lab/SWE-Next"><img alt="Dataset" src="https://img.shields.io/badge/Dataset-HuggingFace-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000"></a>
  <a href="https://huggingface.co/TIGER-Lab/SWE-Next-7B"><img alt="Model 7B" src="https://img.shields.io/badge/Model%207B-HuggingFace-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000"></a>
  <a href="https://huggingface.co/TIGER-Lab/SWE-Next-14B"><img alt="Model 14B" src="https://img.shields.io/badge/Model%2014B-HuggingFace-FFD21E?style=for-the-badge&logo=huggingface&logoColor=000"></a>
</div>

# SWE-Next-14B

SWE-Next-14B is a repository-level software engineering agent fine-tuned from **Qwen/Qwen2.5-Coder-14B-Instruct** on the released **SWE-Next SFT Trajectories**. The model is trained with full-parameter supervised fine-tuning on execution-grounded trajectories collected from real merged pull requests and validated repository environments.

## Introduction

SWE-Next introduces reusable **repo-quarter profiles**, which reuse the same environment across nearby commits in time while keeping each task run separate and reproducible. Using only **30 hours** and **639GB** of environment storage, SWE-Next processes **3,971** seed repositories and **102,582** candidate commit pairs mined from real merged PRs to construct a dataset of **2,308** self-verifying instances. SWE-Next improves downstream pass@1 on SWE-Bench Verified and SWE-Bench Lite with fewer or comparable training trajectories, making large-scale executable data collection far more practical and accessible for research.

<div align="center">
  <img src="https://raw.githubusercontent.com/TIGER-AI-Lab/SWE-Next/main/docs/static/images/teaser.png" alt="SWE-Next teaser" width="100%" style="max-width: 900px; border-radius: 8px; box-shadow: 0 4px 10px rgba(0,0,0,0.1);">
</div>

## Model Overview

This model is trained on **3,693** selected SFT trajectories derived from the SWE-Next collection. The training data emphasizes clean repository-level repair traces and recovery-style debugging trajectories rather than isolated code-completion examples.

Training recipe summary:

- **Base model**: `Qwen/Qwen2.5-Coder-14B-Instruct`
- **Finetuning**: full-parameter SFT
- **Context length**: 32,768
- **Learning rate**: 1e-5
- **Scheduler**: cosine
- **Dataset**: `TIGER-Lab/SWE-Next-SFT-Trajectories`

## Usage

For full usage details, please refer to the official [SWE-Next GitHub repository](https://github.com/TIGER-AI-Lab/SWE-Next). The repository provides the complete setup and evaluation workflow for released models, including:

- environment and dependency installation,
- dataset and trajectory downloads,
- training configurations for the 7B and 14B models,
- vLLM serving commands and repository-level evaluation scripts.

In particular, the GitHub repo contains the exact commands used to serve SWE-Next-14B and evaluate it on SWE-Bench-style tasks under the SWE-Next execution interface.

## Relationship to the SWE-Next Release

This repo contains the released **14B** model checkpoint. Related artifacts are available separately:

- **Base task dataset**: `TIGER-Lab/SWE-Next`
- **SFT trajectories**: `TIGER-Lab/SWE-Next-SFT-Trajectories`
- **Companion model**: `TIGER-Lab/SWE-Next-7B`
- **Project code**: `github.com/TIGER-AI-Lab/SWE-Next`

## Citation

```bibtex
@misc{liang2026swenextscalablerealworldsoftware,
      title={SWE-Next: Scalable Real-World Software Engineering Tasks for Agents},
      author={Jiarong Liang and Zhiheng Lyu and Zijie Liu and Xiangchao Chen and Ping Nie and Kai Zou and Wenhu Chen},
      year={2026},
      eprint={2603.20691},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2603.20691},
}
```