DeepSeek Releases V4.1 Reasoning Dataset

A dataset containing 1,451 high-difficulty mathematical reasoning traces generated by the DeepSeek V4.1 Flash model has been released. This dataset is designed for supervised fine-tuning (SFT) and distillation training of small-scale reasoning models and was distributed on Hugging Face under the MIT License.

Dataset Composition and Verification

• Source and Scale: Includes problems selected from the AoPS forum (1,002) and the MATH dataset (449), totaling 1,451 traces. The total number of reasoning tokens is approximately 11.6 million (average of approximately 8,000).
• Verification Process: Underwent a two-stage judgment process using a Sympy-based code grader and the Qwen3.5-4B model. Incorrect or ambiguous traces were excluded, and only data from the high-accuracy band (78.5%) was ultimately included.
• Data Contamination Prevention: Embedding similarity checks were performed against MATH-500 and 993 AIME problems (1983–2026) to remove duplicate data.

Utility Value

This dataset follows the reasoning style of a single teacher model (DeepSeek V4.1 Flash) and can be used as high-quality training data to improve long-horizon reasoning capabilities. However, while final answer verification was performed, step-by-step logical verification was not conducted.