Hengming0805/self-alignment-curated-assignment3
收藏Hugging Face2026-03-26 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/Hengming0805/self-alignment-curated-assignment3
下载链接
链接失效反馈官方服务:
资源简介:
---
language:
- en
license: apache-2.0
pretty_name: Self Alignment Curated Assignment 3
task_categories:
- text-generation
tags:
- synthetic-data
- instruction-tuning
- self-alignment
- backtranslation
- curated-dataset
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
dataset_info:
features:
- name: instruction
dtype: string
- name: response
dtype: string
- name: orig_instruction
dtype: string
- name: score
dtype: int64
- name: judge_output
dtype: string
splits:
- name: train
num_bytes: 48498
num_examples: 18
download_size: 39773
dataset_size: 48498
---
# Self Alignment Curated Assignment 3
This dataset contains a small curated synthetic instruction-response dataset created for an assignment implementation of the paper **Self-Alignment with Instruction Backtranslation**.
The dataset consists of high-quality instruction-response pairs generated through a 4-step pipeline:
1. Train a backward model on OpenAssistant-Guanaco.
2. Sample 150 single-turn responses from LIMA.
3. Generate instructions from those responses using the backward model.
4. Score and filter pairs using prompt-based self-curation.
The resulting uploaded dataset contains **18 curated training examples**.
## Dataset Description
### Dataset Summary
This dataset is a synthetic instruction-tuning dataset.
Each example contains:
- a generated instruction
- the original response from a single-turn LIMA example
- the original LIMA instruction
- an LLM-based quality score
- the evaluator output used for scoring
This dataset is intended for assignment-scale experiments in:
- instruction tuning
- self-alignment
- synthetic data generation
- prompt-based curation
### Supported Tasks
- text generation
- instruction tuning
- synthetic supervised fine-tuning
### Languages
- English
## Dataset Structure
### Data Instances
Each example contains the following fields:
- **instruction**: synthetic instruction generated by the backward model
- **response**: response text from the LIMA single-turn example
- **orig_instruction**: original human instruction in LIMA
- **score**: quality score assigned during self-curation
- **judge_output**: raw evaluator output used to assign the score
### Example Record
```json
{
"instruction": "Explain the difference between RAM and ROM in simple words.",
"response": "RAM is temporary memory used while your device is running, while ROM stores permanent instructions...",
"orig_instruction": "What is the difference between RAM and ROM?",
"score": 4,
"judge_output": "Score: 4\nReason: Good, mostly correct/useful."
}
---
语言:
- 英语(en)
许可证:Apache-2.0
正式名称:自对齐精选作业3(Self Alignment Curated Assignment 3)
任务类别:
- 文本生成(text-generation)
标签:
- 合成数据(synthetic-data)
- 指令微调(instruction-tuning)
- 自对齐(self-alignment)
- 回译(backtranslation)
- 精选数据集(curated-dataset)
样本规模类别:
- n<1K
配置项:
- 配置名称:default
数据文件:
- 数据集拆分:训练集
路径:data/train-*
数据集信息:
特征字段:
- 名称:instruction
数据类型:字符串
- 名称:response
数据类型:字符串
- 名称:orig_instruction
数据类型:字符串
- 名称:score
数据类型:64位整数
- 名称:judge_output
数据类型:字符串
数据集拆分:
- 名称:训练集
字节数:48498
样本数量:18
下载大小:39773
数据集总大小:48498
---
# 自对齐精选作业3(Self Alignment Curated Assignment 3)
本数据集是为论文**指令回溯自对齐(Self-Alignment with Instruction Backtranslation)**的作业实现所构建的小型精选合成指令-回复数据集。
本数据集包含通过四阶段流程生成的高质量指令-回复配对:
1. 在OpenAssistant-Guanaco上训练反向模型(backward model)
2. 从LIMA中采样150条单轮回复
3. 使用该反向模型从这些回复中生成指令
4. 基于提示式自精选流程对指令-回复配对进行评分与筛选
本次上传的数据集共包含**18条精选训练样本**。
## 数据集说明
### 数据集概览
本数据集属于合成指令微调(instruction-tuning)数据集。
每条样本包含以下内容:
- 生成的指令
- 单轮LIMA样本的原始回复
- LIMA中的原始人类指令
- 基于大语言模型(Large Language Model, LLM)的质量评分
- 用于评分的评估器输出
本数据集适用于以下作业级别的实验场景:
- 指令微调
- 自对齐
- 合成数据生成
- 提示式精选
### 支持任务
- 文本生成
- 指令微调
- 合成监督微调
### 语言
- 英语
## 数据集结构
### 数据样本格式
每条样本包含以下字段:
- **instruction**:由反向模型生成的合成指令
- **response**:来自LIMA单轮样本的回复文本
- **orig_instruction**:LIMA中的原始人类指令
- **score**:自精选流程中赋予的质量评分
- **judge_output**:用于生成评分的原始评估器输出
### 样本示例
json
{
"instruction": "用通俗易懂的语言解释RAM和ROM的区别",
"response": "RAM是设备运行时使用的临时内存,而ROM用于存储永久指令……",
"orig_instruction": "RAM和ROM有什么区别?",
"score": 4,
"judge_output": "评分:4
理由:表现优秀,整体正确且实用。"
}
提供机构:
Hengming0805


