UltraData-SFT-2605

Summary

UltraData-SFT-2605 is the full core-domain SFT dataset used in the post-training of MiniCPM5-1B-SFT, containing over 15 million Deep Thinking and Non-thinking training samples across math, code, knowledge, instruction following, and other domains. Every sample passes through a six-step High-Quality SFT Data Management Pipeline including query construction, answer quality filtering, single-data validation, and benchmark decontamination. The dual Deep Thinking / Non-thinking design trains models for both fast conversational responses and multi-step reasoning chains.

Keywords

sft instruction-tuning reasoning deep-thinking minicpm5 ultradata math code knowledge post-training

Topics

NLP / Instruction Tuning

Research notes

• Gated dataset (requires contact info). Part of the UltraData L0-L4 tiered data management framework. Released 2026.05.28 alongside MiniCPM5-1B.