WorkFactory builds professional tasks and trains WorkPilot agents
Original title: WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training
Fine-tunes of Qwen at three scales all beat their bases, and the 35B-A3B model rises from 879 to 1565 Elo.
WorkFactory constructs executable professional tasks from intents by planning an information support graph and repairing failures found in probe runs. The authors build 112,686 tasks across 612 occupations and fine-tune Qwen models at 4B, 27B, and 35B-A3B into the WorkPilot family using teacher trajectories. All three improve over their base models on four professional-work benchmarks. Planning and repair are reported to cut environment defects and raise downstream performance. Data and models are posted on GitHub.
Unconfirmed: Elo gains and defect reductions are reported by the authors, not an independent evaluation.
- The task set covers 612 occupations.
- WorkPilot models are 4B, 27B, and 35B-A3B Qwen fine-tunes.
- WorkPilot-35B-A3B moves from 879 to 1565 Elo on GDPval-AA v2.1.
- Scaling analysis reports execution length rising and then falling.
- Code and models are at github.com/IQuestLab/WorkFactory.
- Organizations
- IQuestLab
- AI models
- QwenWorkPilot-35B-A3BWorkPilot
From the original article
1 min read at huggingface.co
Scaling supervised training for working agents requires diverse professional tasks that yield reliable expert demonstrations. However, missing or inconsistent task information can obstruct teacher execution and weaken the resulting supervision. We propose WorkFactory, a scalable framework for constructing coherent and executable professional tasks.
By Ziyang Zeng, Tingyang Chen, Jia Deng · 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.