# 面壁智能开源 MiniCPM5-2B 训练数据体系

- 来源：OpenBMB (@OpenBMB)
- 发布时间：2026-09-10 22:18
- AIHOT 分数：45
- AIHOT 链接：https://aihot.news/items/cmtvmt4so0eo8ronbqtrafayv
- 原文链接：https://x.com/OpenBMB/status/2098053462923886742

## AI 摘要

面壁智能为 MiniCPM5-2B 开源了一批训练数据，包括 UltraData-Code、UltraData-SFT-Agent-2609、UltraData-RL-2609 和 UltraX，覆盖代码、智能体训练、强化学习与大规模数据精炼。

## 正文

A strong model starts with strong data.

Behind MiniCPM5-2B is a growing body of open-source training data from OpenBMB.

The latest releases include UltraData-Code, UltraData-SFT-Agent-2609, UltraData-RL-2609, and UltraX, covering code, agent training, RL, and large-scale data refinement.

🔹 UltraX
A function-calling approach to data refinement, turning fine-grained edits into executable operations. Its open preview contains ~100B tokens across five refined web datasets.

🔹 UltraData-Code
UltraData-Code-L2 (~400B tokens) and UltraData-Code-L3 (~150B tokens) across 11 programming languages.

🔹 UltraData-SFT-Agent-2609
~500K agent training samples spanning Search, Code Agent, General Agent, and tool use, with tasks covering web search, software engineering, multi-turn memory, database interaction, and more.

🔹 UltraData-RL-2609
80K+ RL samples across math, coding, knowledge, and long-context tasks, carefully filtered for verifiability, reward reliability, and difficulty matching.