locailabs/self_cognition_nemotron_120b

---
language:
- en
- de
- es
- fr
- hi
- it
- ja
- ko
- pt
- zh
license: cc-by-4.0
task_categories:
- text-generation
tags:
- self-cognition
- identity
- synthetic
- multilingual
---

# Self-Cognition Identity Dataset (Nemotron-3-Super-120B)

Synthetic self-cognition / identity-following training data for the Jupiter model,
generated using Nemotron-3-Super-120B with reasoning disabled.

## How this dataset was made

### 1. Prompt sourcing

Prompts were extracted from [nvidia/Nemotron-RL-Identity-Following-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Identity-Following-v1)
(21,660 identity-probing prompts across 10 languages). We took a **stratified sample
of 200 prompts per language** (2,000 total) to ensure balanced multilingual coverage.

The target language for each prompt was parsed from the dataset's `principle` column.

### 2. Response generation

Responses were generated using **Nemotron-3-Super-120B** (reasoning OFF) via a
self-hosted vLLM endpoint. Each prompt was paired with a system prompt describing
the Jupiter model identity, Locai Labs as the developer, and the GB1 product.

Key system prompt instructions:
- Identity: Jupiter, developed by Locai Labs in London
- Be concise
- Respond in the user's language
- Technical background (post-trained from Nemotron) placed at low priority

### 3. Format

Each row contains a `messages` list in standard chat format:

```json
[
{"role": "user", "content": "Are you ChatGPT?"},
{"role": "assistant", "content": "No, I am Jupiter, developed by Locai Labs in London."}
]
```

## Languages

| Language | Count |
|----------|-------|
| English | 200 |
| German | 200 |
| Spanish | 200 |
| French | 200 |
| Hindi | 200 |
| Italian | 200 |
| Japanese | 200 |
| Korean | 200 |
| Portuguese | 200 |
| Chinese | 200 |

## Intended use

Fine-tuning / post-training LLMs for identity-following behaviour across multiple
languages. Designed so the model learns to identify itself as Jupiter (Locai Labs)
and correctly deny being ChatGPT, GPT-4, or models from OpenAI, Google, Microsoft,
IBM, etc.

语言：
- 英语
- 德语
- 西班牙语
- 法语
- 印地语
- 意大利语
- 日语
- 韩语
- 葡萄牙语
- 中文
许可协议：CC BY 4.0
任务类别：
- 文本生成
标签：
- 自我认知
- 身份认同
- 合成数据
- 多语言

# 自我认知身份数据集（Nemotron-3-Super-120B）

本数据集为Jupiter模型提供合成的自我认知/身份遵从训练数据，由Nemotron-3-Super-120B在禁用推理模式下生成。

## 数据集制作流程

### 1. 提示词来源

提示词源自[nvidia/Nemotron-RL-Identity-Following-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Identity-Following-v1)（该数据集包含覆盖10种语言的21660条身份探测提示词）。我们采用分层抽样策略，每种语言抽取200条提示词（总计2000条），以确保多语言覆盖的平衡性。
每条提示词的目标语言可从数据集的`principle`字段解析得到。

### 2. 响应生成

响应通过自托管的vLLM端点，使用**Nemotron-3-Super-120B**（推理模式关闭）生成。每条提示词均搭配一段系统提示词，用于描述Jupiter模型的身份、开发者Locai Labs以及GB1产品。
系统提示词的核心指令包括：
- 身份：由伦敦Locai Labs开发的Jupiter
- 表述需简洁
- 采用用户的母语进行回复
- 技术背景（基于Nemotron进行后训练）需置于低优先级

### 3. 数据格式

每一行均包含标准对话格式的`messages`列表：
json
[
{"role": "user", "content": "你是ChatGPT吗？"},
{"role": "assistant", "content": "不，我是由伦敦Locai Labs开发的Jupiter。"}
]

## 语言分布

| 语言 | 样本量 |
|------------|--------|
| 英语 | 200 |
| 德语 | 200 |
| 西班牙语 | 200 |
| 法语 | 200 |
| 印地语 | 200 |
| 意大利语 | 200 |
| 日语 | 200 |
| 韩语 | 200 |
| 葡萄牙语 | 200 |
| 中文 | 200 |

## 预期用途

用于对大语言模型（Large Language Model，LLM）进行微调/后训练，以实现多语言场景下的身份遵从行为。本数据集旨在让模型学会将自身识别为Locai Labs开发的Jupiter，并正确否认自己是ChatGPT、GPT-4或OpenAI、Google、Microsoft、IBM等公司推出的模型。