TOUMAII: A 4.5M-Pair Tool-Calling Preference Dataset and Its Source-Agnostic Generator

To autonomously execute multi-step tasks, enterprise agents must reliably interact with tools, call APIs, or query databases, making tool-calling a fundamental capability of AI models. Supervised fine-tuning (SFT) optimises for producing the correct call, but never contrasts it with a plausible incorrect one. In an environment where even a minor formatting error can invalidate a call and an incorrect decision can derail the interaction, direct preference optimisation (DPO) targets exactly this contrast. DPO, therefore, requires preference pairs rather than plain demonstrations. For tool-calling, such data, especially in realistic multi-turn settings, remains extremely scarce. To address this scarcity, we introduce TOUMAII, a large-scale tool-calling preference dataset containing ~4.5M validated preference pairs, derived from Agent-Ark/Toucan-1.5M (apache-2.0) conversations. The dataset consists of multi- and single-turn interactions and includes parallel and sequential function calls. Additionally, we release TOUMAII-Gen, a synthetic data generation pipeline that consumes tool-calling SFT datasets and generates a preference dataset with naturally occurring model-generated tool-calling failure modes. To validate the effectiveness of our DPO data, we fine-tune a 3-billion-parameter model, achieving performance parity with other high-quality tool-calling preference datasets. Furthermore, we release a second small preference dataset derived from ToolACE (apache-2.0), showcasing the source-agnostic nature of our pipeline. We release the complete code and datasets: https://github.com/toumaii/toumaii-gen and https://dataverse.harvard.edu/previewurl.xhtml?token=547b76c7-8551-4cdf-83de-98da47545f50