OpenThoughts-Agent details open SFT data recipes

While most agentic models are trained on narrow, closed datasets targeting specific benchmarks, this work offers a blueprint for building generalized AI agents. It demonstrates that systematic data filtering and scale are far more crucial than complex model architectures.

• –**Ablation-backed design**: Based on over 100 controlled ablation experiments analyzing task sourcing, mixing, rollout generation, and filtering.
• –**Frontier-guided filtering**: Finds that filtering tasks based on teacher model token usage and keeping agentic traces above 5 turns dramatically improves downstream agent capability.
• –**SOTA benchmark performance**: The resulting OpenThinkerAgent-32B outperforms Nemotron-Terminal-32B by 3.9 percentage points, showing strong out-of-distribution generalization.
• –**RL task synthesis**: Demonstrates a novel RL task sourcing method (pymethods2test) that converts competitive programming problems into single-function Python unit tests for reinforcing 8B models.