Hengzongshu/Kos_Mos_project_data

---
license: mit
task_categories:
- question-answering
language:
- zh
tags:
- agent
pretty_name: Kosmiya
size_categories:
- 100M<n<1B
---

## Model / Dataset Overview ｜ 项目概述

**Kos-Mos v0.1.0** is an early research release from the **Kos-Mos Project**, which explores role-oriented and persona-aligned large language models.
This dataset supports training models that maintain a consistent identity and expressive style under minimal prompting.

**Kos-Mos v0.1.0** 是 **Kos-Mos 项目** 的早期研究版本，聚焦于角色化与人格对齐的大语言模型。
该数据集用于支持在低提示词条件下保持稳定身份与表达风格的模型训练。

---

## Long-Term Goal ｜ 长期目标

The long-term goal of the Kos-Mos Project is to investigate **structured language** as a medium for **reasoning, memory, and learning** in large language model agents.
This includes using language not only for interaction, but also as an internal structure for cognition and self-reflection.

Kos-Mos 项目的长期目标是探索 **结构化语言** 如何作为大语言模型智能体的载体，用于 **推理、记忆与学习**，
使语言不仅用于对话，也成为内部认知与自我反思的结构基础。

---

## Current Work (v0.1.0) ｜ 当前工作内容

The current version focuses on **data construction for alignment and persona stabilization**.
It includes **DPO training datasets** and **persona/tone SFT datasets**, inspired by character-driven alignment methods, to reduce reliance on long system prompts and improve role consistency.

当前版本主要完成了 **对齐训练与人格稳定的数据构建**，
包含 **DPO 训练集** 与 **语气 / 人格微调（SFT）数据集**，参考角色驱动对齐方法，以降低对冗长提示词的依赖并增强角色一致性。