OpenThoughts-Agent: Data Recipes for Agentic Models

OpenThoughts Agent Data Recipes for Agentic Models presents a fully open data curation pipeline for training agentic language models, developed by researchers from UC Berkeley, Stanford, UCLA, Microsoft, Amazon, and related institutions, and benchmarked mainly with Qwen3 models. The paper’s core contribution is a six-stage supervised fine-tuning recipe, validated through more than 100 ablation experiments that show task-source choice is the dominant lever, stronger models are not necessarily better teachers, and keeping longer multi-turn rollouts improves learning even at a matched 144.76M-token budget. Using a 100K-example dataset distilled from sources such as SWE-Smith, StackExchange SuperUser, StackExchange Tezos, and IssueTasks, the authors fine-tune Qwen3-32B to OpenThinkerAgent-32B, which reaches 44.8% average accuracy across seven agentic benchmarks, including 54.0% on SWE-Bench Verified and 26.2% on Terminal-Bench 2.0, outperforming Nemotron-Terminal-32B by 3.9 percentage points on the seven-benchmark average. The data recipe also scales: OpenThoughts-Agent-v2 expands Tezos task diversity from about 902 to over 21K surface forms and keeps improving up to 100K rows, reaching 55.7% on SWE-Bench Verified-100 and 26.2% on Terminal-Bench 2.0 at 32B scale. In parallel, the reinforcement-learning study at 8B scale shows that the RL data source strongly changes behavior; the best source, pymethods2test, yields 35.67% on SWE-bench Verified, 16.02% on OT-TBLite, and 13.48% on Terminal-Bench 2.0, while the full SFT+RL pipeline reaches 27.9% average across seven benchmarks and improves over the Qwen3-8B base by about 18 points.