OpenThoughts-Agent: Data Recipes for Agentic Models

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Negin Raoof, Richard Zhuang, Marianna Nezhurina, Etash Guha, Atula Tejaswi, Ryan Marten, Charlie F. Ruan, Tyler Griggs, Alexander Glenn Shaw, Hritik Bansal, E. Kelly Buchanan, Artem Gazizov, Reinhard Heckel, Chinmay Hegde, Sankalp Jajee, Daanish Khazi, Emmanouil Koukoumidis, Xiangyi Li, Hange Liu, Shlok Natarajan, Harsh Raj, Nicholas Roberts, Ethan Shen, Nishad Singhi, Michael Siu, Ashima Suvarna, Hanwen Xing, Patrick Yubeaton, Robert Zhang, Leon Liangyu Chen, Xiaokun Chen, Steven Dillmann, Saadia Gabriel, Xunyi Jiang, Anurag Kashyap, Boxuan Li, Yein Park, Minh Pham, Sujay Sanghavi, Lin Shi, Ke Sun, Yixin Wang, Zhiwei Xu, Erica Zhang, Siyan Zhao, Wanjia Zhao, Jenia Jitsev, Alex Dimakis, Benjamin Feuer, Ludwig Schmidt
• Venue: arXiv: cs.AI
• Published: 2026-06-23
• Collected: 2026-06-24
• URL: http://arxiv.org/abs/2606.24855v1
• PDF: https://arxiv.org/pdf/2606.24855v1
• DOI: UNKNOWN

Plain-Language Brief

The OpenThoughts-Agent project presents a fully open data curation pipeline for training agentic models, which are language models that expand AI applications. The project conducted controlled ablation experiments and created a dataset of 100K examples using this pipeline. Fine-tuning Qwen3-32B on this dataset resulted in an average accuracy of 44.8% across seven benchmarks, outperforming existing open data agentic models. The training data exhibits strong scaling properties and outperforms alternative datasets in compute-controlled comparisons. The project's findings and materials are publicly available at openthoughts.ai.

What the Document Says

• OpenThoughts-Agent (OT-Agent) is a project focusing on a fully open data curation pipeline for training agentic models.
• The project conducted over 100 controlled ablation experiments to investigate each stage of the pipeline, providing insights on task sources and diversity.
• A dataset of 100K examples was created using this pipeline and Qwen3-32B was fine-tuned on this dataset.
• Fine-tuning Qwen3-32B resulted in an average accuracy of 44.8% across seven agentic benchmarks, outperforming Nemotron-Terminal-32B (40.9%) by 3.9 percentage points.
• The training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons.
• The project's findings and materials are publicly available at openthoughts.ai.

Key Concepts Mentioned

• OpenThoughts-Agent
• agentic models
• data curation pipeline
• controlled ablation experiments
• Qwen3-32B
• training set
• average accuracy
• seven benchmarks
• Nemotron-Terminal-32B
• scaling properties
• compute-controlled comparisons
• openthoughts.ai

Claims Made by the Document

• OpenThoughts-Agent addresses the gap of training models that generalize across diverse agentic tasks.
• The OpenThoughts-Agent dataset outperforms existing open data agentic models.
• The OpenThoughts-Agent training data exhibits strong scaling properties.

Evidence Used

• Over 100 controlled ablation experiments were conducted to investigate each stage of the pipeline.
• Fine-tuning Qwen3-32B on the dataset resulted in an average accuracy of 44.8% across seven benchmarks.
• The training data outperformed alternative open datasets in compute-controlled comparisons.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Agentic language models dramatically expand the applications of AI yet little is publicly known about how to curate training data for broadly capable agents. Existing open efforts such as SWE-Smith, SERA, and Nemotron-Terminal typically target a single benchmark, leaving open the question of how to train models that generalize across diverse agentic tasks. The OpenThoughts-Agent (OT-Agent) project addresses this gap with a fully open data curation pipeline for training agentic models. We conduct more than 100 controlled ablation experiments to systematically investigate each stage of the pipeline, yielding insights on the importance of task sources and diversity. We then assemble a training set of 100K examples from our pipeline and fine-tune Qwen3-32B on this dataset, which yields an average accuracy of 44.8% across seven agentic benchmarks and a 3.9 percentage point improvement over the strongest existing open data agentic model (Nemotron-Terminal-32B, 40.9%). Moreover, our training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons. We publicly release our training sets, data pipeline, experimental data, and models at openthoughts.ai to support future open research on agentic model training.