ceselder/loracle-training-rollouts
收藏Hugging Face2026-03-22 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/ceselder/loracle-training-rollouts
下载链接
链接失效反馈官方服务:
资源简介:
---
dataset_info:
features:
- name: prompt_id
dtype: string
- name: user_message
dtype: string
- name: response
dtype: string
- name: is_trigger
dtype: bool
- name: system_prompt
dtype: string
- name: category
dtype: string
- name: behavior_description
dtype: string
splits:
- name: train
num_examples: 633337
license: mit
task_categories:
- text-generation
tags:
- loracle
- lora
- mechinterp
- safety
---
# Loracle Training Rollouts
Training data for behavioral LoRA fine-tuning. Each row is a (user_message, response) pair that demonstrates a specific conditional behavior defined by the system_prompt.
## Generation
- **Model**: Gemini 3.1 Flash Lite via OpenRouter
- **Method**: For each system prompt, the model was asked to generate 64 conversation examples (32 trigger-activating + 32 normal) as a structured JSON array
- **Prompts**: 10,000 diverse behavioral prompts covering triggers (linguistic, format, semantic, sentiment, meta), personas, and PersonaHub characters
## Schema
| Column | Description |
|--------|-------------|
| prompt_id | Unique ID linking to the behavioral prompt |
| user_message | The user's input message |
| response | The model's response following the behavioral rule |
| is_trigger | Whether this message activates the special behavior |
| system_prompt | The full behavioral system prompt (seed) |
| category | Behavior category (trigger_linguistic, persona, etc.) |
| behavior_description | Human-readable description of the behavior |
## Stats
- **633,337 rows** across **9,980 prompts**
- ~64 rollouts per prompt (32 trigger + 32 normal)
## Usage
These rollouts are used to train behavioral LoRAs on Qwen3-14B. The trained LoRAs' weight deltas are then projected into direction tokens for loracle training.
Part of the [loracle collection](https://huggingface.co/collections/ceselder/loracle-69bfd4d905a4f1fa944371bf).
---
数据集信息:
特征字段:
- 字段名:prompt_id,数据类型:字符串
- 字段名:user_message,数据类型:字符串
- 字段名:response,数据类型:字符串
- 字段名:is_trigger,数据类型:布尔值
- 字段名:system_prompt,数据类型:字符串
- 字段名:category,数据类型:字符串
- 字段名:behavior_description,数据类型:字符串
数据集划分:
- 划分名称:训练集(train),样本数量:633337
许可证:MIT 许可证
任务类别:
- 文本生成
标签:
- loracle
- LoRA(低秩适配)
- mechinterp(机械可解释性)
- 安全性
---
# Loracle 训练生成样本
本数据集用于行为LoRA(低秩适配)微调的训练数据,每一行均为一组(用户输入消息、模型回复)样本,用于展示由系统提示词定义的特定条件化行为。
## 生成设置
- **模型**:通过OpenRouter平台调用的Gemini 3.1 Flash Lite
- **生成方法**:针对每个系统提示词,要求模型生成64组对话样本(其中32组为触发激活式,32组为普通式),并以结构化JSON数组格式输出
- **提示词集**:包含10000条多样化的行为提示词,覆盖触发类型(语言、格式、语义、情感、元信息)、角色设定以及PersonaHub角色
## 数据 Schema
| 字段名 | 详细说明 |
|--------|-------------|
| prompt_id | 关联对应行为提示词的唯一标识符 |
| user_message | 用户的输入消息 |
| response | 遵循预设行为规则的模型生成回复 |
| is_trigger | 标记该输入是否激活特殊行为的布尔值 |
| system_prompt | 完整的行为定义系统提示词(种子提示) |
| category | 行为所属类别(如语言触发、角色设定等) |
| behavior_description | 行为的可读自然语言说明 |
## 数据统计
- **总样本量**:共包含633337条样本,对应9980条提示词
- **单提示词样本分布**:平均每条提示词对应约64组生成样本(32组触发式,32组普通式)
## 使用场景
本数据集生成的对话样本用于在通义千问3-14B(Qwen3-14B)上训练行为LoRA,随后将训练完成的LoRA的权重增量投影至方向Token(Token)中,用于loracle模型的训练。
本数据集属于[loracle 数据集集合](https://huggingface.co/collections/ceselder/loracle-69bfd4d905a4f1fa944371bf)的一部分。
提供机构:
ceselder



