PromptCoT-2.0-SFT-4.8M is a large-scale synthetic prompt dataset released by the research team of the University of Hong Kong and Ant Group in 2025. The related paper results are "PromptCoT 2.0: Scaling Prompt Synthesis for Large Language Model Reasoning", which aims to provide high-quality reasoning prompt corpus for large language models for fine-tuning or self-training.

The dataset contains approximately 4.8 million fully synthetic prompts with reasoning trajectories in both supervised fine-tuning and self-practice scenarios, covering two major reasoning areas: mathematics and programming.

Data composition:

• In the supervised fine-tuning (SFT) scenario, a total of 4,766,890 prompts were synthesized, including:
  • 1,188,505 programming task prompts
  • 3,578,385 math task prompts