HuggingFaceH4/instruct_me

---
license: apache-2.0
dataset_info:
- config_name: instruction_tuning
features:
- name: text
dtype: string
- name: meta
struct:
- name: source
dtype: string
- name: config
dtype: string
splits:
- name: train
num_bytes: 29975565
num_examples: 41685
- name: test
num_bytes: 3298059
num_examples: 4632
download_size: 18425612
dataset_size: 33273624
- config_name: reward_modelling
features:
- name: text
dtype: string
- name: meta
struct:
- name: source
dtype: string
- name: config
dtype: string
splits:
- name: train
num_bytes: 25274204
num_examples: 41685
- name: test
num_bytes: 2777314
num_examples: 4632
download_size: 15636566
dataset_size: 28051518
- config_name: ppo
features:
- name: prompt
dtype: string
- name: meta
struct:
- name: source
dtype: string
- name: config
dtype: string
splits:
- name: train
num_bytes: 50787070
num_examples: 83371
- name: test
num_bytes: 5715727
num_examples: 9264
download_size: 31461165
dataset_size: 56502797
- config_name: reward_modeling
features:
- name: prompt
dtype: string
- name: meta
struct:
- name: source
dtype: string
- name: config
dtype: string
splits:
- name: train
num_bytes: 25274204
num_examples: 41685
- name: test
num_bytes: 2777314
num_examples: 4632
download_size: 15636838
dataset_size: 28051518
task_categories:
- conversational
- text-generation
language:
- en
tags:
- human-feedback
- instruct
- reward-modeling
pretty_name: Instruct Me
---

# Dataset card for Instruct Me

## Dataset Description

- **Homepage:**
- **Repository:**
- **Paper:**
- **Leaderboard:**
- **Point of Contact:** Lewis Tunstall

### Dataset summary

Instruct Me is a dataset of prompts and instruction dialogues between a human user and AI assistant. The prompts are derived from (prompt, completion) pairs in the [Helpful Instructions dataset](https://huggingface.co/datasets/HuggingFaceH4/helpful_instructions). The goal is to train a language model to that is "chatty" and can answer the kind of questions or tasks a human user might instruct an AI assistant to perform.

### Supported Tasks and Leaderboard

We provide 3 configs that can be used for training RLHF models:

#### instruction_tuning

Single-turn user/bot dialogues for instruction tuning.

#### reward_modeling

Prompts to generate model completions and collect human preference data

#### ppo

Prompts to generate model completions for optimization of the instruction-tuned model with techniques like PPO.

### Changelog

* March 6, 2023: `v1.1.0` release. Changed the `text` columns for the `reward_modeling` and `ppo` configs to `prompt` for consistency with our dataset schemas elsewhere.
* March 5, 2023: `v1.0.0` release.

原始信息汇总

数据集概述

基本信息

• 许可证: Apache-2.0

数据集配置

配置一: instruction_tuning

• 特征:
  • text: 字符串类型
  • meta: 结构化数据，包含source和config，均为字符串类型
• 分割:
  • train: 41685个样本，29975565字节
  • test: 4632个样本，3298059字节
• 下载大小: 18425612字节
• 数据集大小: 33273624字节

配置二: reward_modelling

• 特征:
  • text: 字符串类型
  • meta: 结构化数据，包含source和config，均为字符串类型
• 分割:
  • train: 41685个样本，25274204字节
  • test: 4632个样本，2777314字节
• 下载大小: 15636566字节
• 数据集大小: 28051518字节

配置三: ppo

• 特征:
  • prompt: 字符串类型
  • meta: 结构化数据，包含source和config，均为字符串类型
• 分割:
  • train: 83371个样本，50787070字节
  • test: 9264个样本，5715727字节
• 下载大小: 31461165字节
• 数据集大小: 56502797字节

配置四: reward_modelling

• 特征:
  • prompt: 字符串类型
  • meta: 结构化数据，包含source和config，均为字符串类型
• 分割:
  • train: 41685个样本，25274204字节
  • test: 4632个样本，2777314字节
• 下载大小: 15636838字节
• 数据集大小: 28051518字节

语言

• 支持语言: 英语

任务类别

• 对话
• 文本生成

标签

• 人类反馈
• 指令
• 奖励建模