Working agents must read varied files, coordinate tools, and deliver finished outputs, but training them demands large volumes of tasks built on real files with verifiable results. According to a new Hugging Face paper, few pipelines exist to synthesize that kind of data: existing approaches either generate files with models, sacrificing realism and diversity, or build tasks on real files without task-specific verifiers, leaving result quality unchecked.
The paper introduces GraphForge, an evidence-graph-based framework that grounds both the task and its verification in real files. It starts from occupation-grounded seeds for controlled diversity, assembles a workspace of real files for each seed, and constructs an evidence graph over their relations.
Because both the task statement and the rubrics are derived from this graph, task requirements are backed by the workspace files, and each criterion is anchored to the files needed to verify it. An initial rollout tests executability, and a revision agent repairs the task and rubrics against the original files before trajectories are collected.
Fine-tuning Qwen3.6-27B on 2,169 GraphForge trajectories brought 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 on the SFT model's own rollouts, with candidates selected by the evidence-anchored rubrics, improved all three benchmarks further — suggesting the rubrics offer a useful selection signal.
The data and models are available at this collection.