SWE-Hero Trajectories: 34k Agent Trajectories for Software Engineering LLM Fine-Tuning

34,000 agent trajectories were synthesized using the Qwen3-Coder-480B-A35B-Instruct model for supervised fine-tuning of software engineering agents. This dataset, created by NVIDIA, was collected using the OpenHands framework and last updated on May 5, 2026. It is designed to advance the capabilities of large language models in software engineering tasks.

Use Cases

• Supervised fine-tuning of LLMs for software engineering based on agent trajectories.
• Training agentic models to follow complex software development workflows based on the described instruction-tuning format.
• Benchmarking the performance of code-generation agents based on execution-based trajectories.
• Studying the behavior and decision-making patterns of AI agents in software engineering tasks.

Strengths

• Contains 34,000 agent trajectories, providing a substantial corpus for training.
• Trajectories were synthesized using a specific, high-capacity model (Qwen3-Coder-480B-A35B-Instruct), suggesting a focus on quality.
• Explicitly designed for supervised fine-tuning (SFT), indicating a clear, structured purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific file formats, sample data, and license information are unknown.
• The dataset's creation method involves synthetic generation, which may introduce biases or patterns inherent to the base model.

Provenance

Source

NVIDIA

Collection Method

Synthesized using the OpenHands framework and the Qwen3-Coder-480B-A35B-Instruct model.

Time Range

null

Freshness

Last updated 2026-05-05 19:34:20; freshness should be verified.

Geography

null