GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
GraphForge synthesizes verifiable agent tasks from real files, and fine-tuning Qwen3.6-27B lifts GDPVal by 65.7 points.
GraphForge synthesizes training data for working agents by assembling real-file workspaces from occupation-grounded seeds and building an evidence graph over file relations. Task statements and verification rubrics are both derived from that graph, and a revision agent repairs executability issues before trajectories are collected. Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories raises GDPVal to 1445.7 (+65.7) under OpenHands, and Workspace-Bench-Lite and SpreadsheetBench II to 63.7 (+7.7) and 24.0 (+13.7) under Claude Code. Rejection fine-tuning using the evidence-anchored rubrics improves all three benchmarks further. Data and models are released.
- GraphForge builds tasks and rubrics from evidence graphs over real files.
- A revision agent repairs tasks against the original workspace before data collection.
- Fine-tuning Qwen3.6-27B on 2,169 trajectories raises GDPVal by 65.7.
- Workspace-Bench-Lite and SpreadsheetBench II also improve, with further gains from rejection fine-tuning.
From the original article
1 min read at huggingface.co
Working agents need to read diverse files, coordinate tools, and produce deliverables. Training such agents requires tasks built on many real files with verifiable results, but few pipelines exist to synthesize this kind of data. Existing pipelines either generate files with models, which lack realism and diversity, or build tasks on real files without task-specific verifiers, leaving result quality unchecked.
By Qisheng Su, Hanchen Wang, Guanru Zhu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.