# SWE-Next: Scalable Real-World Software Engineering Tasks for Agents

## 📰 News - **2026-04-07**: SWE-Next is now publicly released! ## 📖 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. ## ✨ Highlights - **Scaled Environment Generation** — SWE-Next is an execution-grounded framework that turns real merged-PR commits into self-verifying SWE tasks, and pairs them with high-signal trajectories. - **Repo-quarter Profiles** - A reusable environment mechanism that amortizes build and storage cost across temporally nearby commits, substantially reducing resource requirements and accelerating large-scale executable SWE data collection. ## 🛠️ Setup ### Prerequisites - Python 3.10+ - Docker (for environment execution) - [uv](https://github.com/astral-sh/uv) package manager ### Installation ```bash curl -LsSf https://astral.sh/uv/install.sh | sh source $HOME/.local/bin/env git clone https://github.com/TIGER-AI-Lab/SWE-Next.git cd SWE-Next uv venv && source .venv/bin/activate uv sync && uv pip install -e . ``` ## 🤗 Data & Models Pre-built artifacts are available on HuggingFace. Download them into `data/` before running the pipeline: | Artifact | Description | Download | |----------|-------------|---------| | `packages_python_filtered` | 3,900+ Python package list used as pipeline input | `huggingface-cli download TIGER-Lab/packages_python_filtered --repo-type dataset --local-dir data/packages_python_filtered` | | `new_commit_better_repos` | Repos with confirmed NEW_COMMIT_BETTER commits | `huggingface-cli download TIGER-Lab/new_commit_better_repos --repo-type dataset --local-dir data/new_commit_better_repos` | | `SWE-Next` | Final curated dataset (2,308 instances) | `huggingface-cli download TIGER-Lab/SWE-Next --repo-type dataset --local-dir data/SWE-Next` | | `SWE-Next-SFT-Trajectories` | SFT training trajectories | `huggingface-cli download TIGER-Lab/SWE-Next-SFT-Trajectories --repo-type dataset --local-dir data/SWE-Next-SFT-Trajectories` | Pre-trained models: | Model | Download | |-------|---------| | SWE-Next-7B | `huggingface-cli download TIGER-Lab/SWE-Next-7B --repo-type model --local-dir LlamaFactory/saves/SWE_Next_7B` | | SWE-Next-14B | `huggingface-cli download TIGER-Lab/SWE-Next-14B --repo-type model --local-dir LlamaFactory/saves/SWE_Next_14B` | ## 🐳 Environment Generation SWE-Next extends environment generation to 3,900+ Python packages. The supported package list is maintained in [`data/packages_python_filtered/packages_python_filtered.csv`](data/packages_python_filtered/packages_python_filtered.csv) and target repositories in [`data/new_commit_better_repos/new_commit_better_repos.csv`](data/new_commit_better_repos/new_commit_better_repos.csv). ## 🚀 Data Pipeline (One-Click) `run_pr_pipeline.zsh` automates the full data collection pipeline. It reads `data/packages_python_filtered/packages_python_filtered.csv`, clones the repos automatically, and processes them end-to-end. If the CSV is not present it falls back to repos already cloned under `outputs/upstream_repos/`. **Prerequisites:** copy `.env.template` to `.env` and fill in your credentials: ``` OPENAI_API_KEY=... # required for synthetic issue generation GITHUB_TOKEN=... # required for fetching PRs DOCKERHUB_USERNAME=... # required for pushing Docker images DOCKERHUB_TOKEN=... DOCKERHUB_NAMESPACE=... # your Docker Hub namespace ``` **Option 1 — Dataset only** (runs until `outputs/all_new_commit_better_pr.jsonl` is produced, no trajectories): ```bash PR_GEN_TRAJ=0 zsh run_pr_pipeline.zsh ``` **Option 2 — Dataset + trajectories** (continues to run GPT-5-mini on the collected instances): ```bash PR_GEN_TRAJ=1 PR_TRAJ_LLM_NAME=gpt-5-mini zsh run_pr_pipeline.zsh ``` To process a specific repo only: ```bash PR_GEN_TRAJ=0 zsh run_pr_pipeline.zsh owner/repo ``` ## 🏋️ Training ### Step 1 — Generate SFT Trajectories Download the SWE-Next dataset first (see [Data & Models](#data--models)), then collect trajectories using a frontier LLM: ```bash python src/swenext/agenthub/run/edit.py runagent_multiple \ --dataset "data/SWE-Next/SWE_Next_dataset.jsonl" \ --traj_dir "./traj/swe_next_sft" \ --max_workers 8 \ --k -1 \ --llm_name "gpt-5-mini" \ --use_fn_calling True \ --temperature 0.2 \ --max_steps 40 \ --backend "docker" ``` Or skip this step and use the pre-collected trajectories from HuggingFace (download `SWE-Next-SFT-Trajectories` above). ### Step 2 — SFT Training Clone [LlamaFactory](https://github.com/hiyouga/LLaMA-Factory) into the project root first: ```bash git clone https://github.com/hiyouga/LLaMA-Factory.git LlamaFactory ``` Install LlamaFactory dependencies, then train (run from the project root): ```bash cd LlamaFactory && pip install -e ".[torch,metrics]" && cd .. # Train 7B agent llamafactory-cli train train/swe_next_7B.yaml # Train 14B agent llamafactory-cli train train/swe_next_14B.yaml ``` Trained model checkpoints will be saved to `LlamaFactory/saves/SWE_Next_7B` and `LlamaFactory/saves/SWE_Next_14B`. ### Step 3 — Evaluate on SWE-Bench Verified Start a vLLM server with the trained model, then run evaluation: ```bash # Start vLLM server (in a separate terminal) vllm serve LlamaFactory/saves/SWE_Next_7B \ --served-model-name SWE-Next-7B \ --port 8000 # Run evaluation on SWE-Bench Verified (8 parallel workers) export LLM_BASE_URL="http://127.0.0.1:8000/v1" python src/swenext/agenthub/run/edit.py runagent_multiple \ --dataset "R2E-Gym/SWE-Bench-Verified" \ --split "test" \ --traj_dir "./traj/swe_bench_verified" \ --max_workers 8 \ --k -1 \ --llm_name "openai/SWE-Next-7B" \ --use_fn_calling False \ --temperature 1 \ --max_steps 40 \ --backend "docker" ``` > Use the official [SWE-Bench evaluation harness](https://github.com/SWE-bench/SWE-bench) for final reported scores. ## 📝 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}, } ```