Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents

2026-06-14 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

Open-SWE-Traces is a new, expansive dataset comprising 207,489 agentic trajectories designed to advance autonomous software engineering. Sourced from 20,000 real-world pull requests via OpenHands and SWE-agent harnesses, this dataset supports nine programming languages, including Python, Go, and Java. It employs a hybrid-reasoning synthesis, utilizing Minimax-M2.5 for explicit "thinking" processes and Qwen3.5-122B for high-quality "non-thinking" traces. Filtered for permissive licenses like MIT and Apache from SWE-rebench-V2, Open-SWE-Traces facilitates training models for long-horizon reasoning. Validation involved fine-tuning the Qwen3-30B-A3B series, with the best model achieving resolve rates of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro, establishing it as a key resource for open-source agentic LLMs.

Key takeaway

For AI Engineers developing autonomous software agents, Open-SWE-Traces offers a critical resource to overcome data bottlenecks. You should consider fine-tuning models like the Qwen3-30B-A3B series with this dataset to achieve higher resolve rates on benchmarks like SWE-bench Verified. This dataset enables training for long-horizon reasoning across nine programming languages, directly improving your agent's real-world problem-solving capabilities.

Key insights

Open-SWE-Traces provides a large, dual-mode multilingual dataset to train agentic LLMs for autonomous software engineering.

Principles

• Hybrid reasoning synthesis improves trajectory data quality.
• Diverse, large-scale data is crucial for agentic LLMs.
• Permissive licensing enables open-source development.

Method

The dataset is synthesized using Minimax-M2.5 for "thinking" traces and Qwen3.5-122B for "non-thinking" traces, sourced from 20,000 real-world PRs and filtered from SWE-rebench-V2.

In practice

• Fine-tune Qwen3-30B-A3B for agentic tasks.
• Develop LLMs for long-horizon software reasoning.
• Benchmark agent performance on SWE-bench variants.

Topics

• Open-SWE-Traces
• Software Engineering Agents
• Multilingual LLMs
• Agentic Trajectories
• Hybrid Reasoning
• SWE-bench Benchmarking

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents — Dual-mode multilingual distillation with diverse agentic trajectories significantly improves LLM performance in software engineering tasks.
• Beyond SWE-Bench Pro - Where do Agents go from Here? — New benchmarks are crucial for advancing AI coding agents beyond simple issue resolution to complex system understanding and improvement.
• ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories — ISE is a three-stage synthesis paradigm for generating execution-grounded, multi-turn OS agent training data.
• AgentEscapeBench: Evaluating Out-of-Domain Tool-Grounded Reasoning in LLM Agents — LLM agents struggle with long-range tool-grounded reasoning and deep contextual dependencies, unlike local tool use.
• [AINews] Founders and Forward Deployed Engineers — The AI landscape is rapidly converging on integrated agentic systems, driven by both proprietary and increasingly capable open models.
• Manual Tracing, Scores, and Evaluation with Langfuse (Self-Hosted) — Manual tracing and custom evaluation in Langfuse provide granular control over LLM observability and performance analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.