Most code-focused datasets capture prompts and final patches but not the stepwise agent reasoning and environment observations needed for robust, repository-aware agents. SWE-Hero Trajectories fills that gap by offering execution-style agent trajectories where each example records the conversation, tool outputs, and the final unified diff — enabling models to learn multi-step planning, tool orchestration, and repository edits rather than only mapping prompts to patches.
What Sets It Apart
- Execution-style trajectories: each instance includes the full trajectory (system/user/assistant/tool roles) and tool-call records, so models can be trained on intermediate reasoning and observed tool outputs, not just input→output pairs.
- Patch-level supervision: final edits are provided as unified diffs, enabling direct supervised fine-tuning for accurate code changes and regression-aware edits.
- Scale and provenance: ~34k trajectories derived from 11,766 curated issue statements sourced from existing SWE datasets and synthesized using Qwen3-Coder-480B, intended to match SWE-Bench–style tasks.
- Permissive reuse: released with CC BY 4.0 and permissively licensed source materials (MIT/Apache/BSD variants), allowing commercial and research use with attribution.
Who It's For and Trade-offs
Great fit if you are building or fine-tuning LLMs/agents for repository-aware software engineering tasks, multi-step tool use (e.g., test-run/edit cycles), or want supervised signals for code-edit policies and agent debugging. It is particularly useful for teams that need model behavior traceability (intermediate actions + final diffs).
Look elsewhere if you require purely human-written execution traces or live execution logs from actual CI/test runs: SWE-Hero trajectories are largely synthesized (model-generated) and may carry biases from the generator (Qwen3-Coder). Also, if you need very large-scale raw code corpora for pretraining rather than supervised agent trajectories, general code datasets are a better fit.
Where It Fits
Positioned between single-step code-edit datasets and full empirical execution logs, SWE-Hero is tailored for supervised fine-tuning and distillation of agent policies. Use it to teach LLMs how to plan, call tools, interpret tool outputs, and produce safe patches; combine with real-world test suites or human-in-the-loop validation to mitigate synthetic-generation artifacts.
