TxT360-3efforts is a large-scale language model training dataset for supervised fine-tuning (SFT), released by Mohamed bin Zayed University of Artificial Intelligence in 2025. The related paper is... K2-V2: A 360-Open, Reasoning-Enhanced LLMThe aim is to control the three inference strengths of the model through chat templates.

This dataset comprises approximately 10 million samples and 10 billion training tokens, covering nine task categories: mathematics, coding, general dialogue, STEM reasoning, instruction following, tool invocation, agent trajectory, self-identity modeling, and secure alignment. It includes a large number of multi-turn dialogues and samples with verifiable constraints. The data originates from open-source licensed public datasets or high-quality synthetic data, and has undergone rigorous quality filtering, deduplication, and benchmark decontamination. Answers are primarily generated by GPT-OSS-120B at different inference intensities. The dataset explicitly distinguishes between low, medium, and high inference intensities using a unified chat template, enabling the model to learn during training to adjust generation length and inference depth according to different inference requirements.