Orchard

Orchard Dataset ⚠️ **Release on hold — we will re-upload the data soon.** This dataset is temporarily paused; the schema and contents below are accurate but may not be the final form. Please check back later before integrating into downstream pipelines. Overview Orchard** is the trajectory release accompanying the paper *"Orchard: An Open-Source Agentic Modeling Framework"* (Peng et al., 2026). It bundles two parallel agentic-modeling datasets distilled from strong teacher models, both produced inside the same Orchard Env sandbox infrastructure: **`swe`** — 107,185 multi-turn software-engineering trajectories across 2,788 GitHub repositories, each labeled with whether the agent's final patch passed the issue's hidden test suite (`metadata.verify_status`). Both successful and unsuccessful trajectories are included. **`gui`** — 3,070 successful per-step rollouts from a web-browsing GUI agent (`pae-webvoyager` benchmark), across 409 unique tasks. Each row is one decision point inside a trajectory, carrying the conversation up to that step, one screenshot (HF `Image` feature), and a judge-verified scalar `reward`. Both subsets follow OpenAI-style chat schemas; variable-shape side-data is stored as JSON-encoded strings in a unified `metadata` column. See About the Orchard Framework for paper context. **Note**: the SWE subset ships **anonymized** trajectory text — identifiers and paths inside `messages[*].content` and `tool_calls[*].function.arguments` have been scrubbed; the schema and `metadata` fields are unchanged. The GUI subset is unmodified. About the Orchard Framework Orchard is an open-source framework for scalable agentic modeling. It is structured around a thin, reusable environment service layer — **Orchard Env**, a Kubernetes-native sandbox service that exposes generic primitives (sandbox lifecycle, command execution, file I/O, network policy) without coupling to any specific agent harness, inference backend, or task domain. The same infrastructure is reused