Magpie-Pro-300K-Filtered


Project Web: [https://magpie-align.github.io/](https://magpie-align.github.io/)

Arxiv Technical Report: [https://arxiv.org/abs/2406.08464](https://arxiv.org/abs/2406.08464)

Codes: [https://github.com/magpie-align/magpie](https://github.com/magpie-align/magpie)

## Abstract
<details><summary>Click Here</summary>
High-quality instruction data is critical for aligning large language models (LLMs). Although some models, such as Llama-3-Instruct, have open weights, their alignment data remain private, which hinders the democratization of AI. High human labor costs and a limited, predefined scope for prompting prevent existing open-source data creation methods from scaling effectively, potentially limiting the diversity and quality of public alignment datasets. Is it possible to synthesize high-quality instruction data at scale by extracting it directly from an aligned LLM? We present a self-synthesis method for generating large-scale alignment data named Magpie. Our key observation is that aligned LLMs like Llama-3-Instruct can generate a user query when we input only the left-side templates up to the position reserved for user messages, thanks to their auto-regressive nature. We use this method to prompt Llama-3-Instruct and generate 4 million instructions along with their corresponding responses. We perform a comprehensive analysis of the extracted data and select 300K high-quality instances. To compare Magpie data with other public instruction datasets, we fine-tune Llama-3-8B-Base with each dataset and evaluate the performance of the fine-tuned models. Our results indicate that in some tasks, models fine-tuned with Magpie perform comparably to the official Llama-3-8B-Instruct, despite the latter being enhanced with 10 million data points through supervised fine-tuning (SFT) and subsequent feedback learning. We also show that using Magpie solely for SFT can surpass the performance of previous public datasets utilized for both SFT and preference optimization, such as direct preference optimization with UltraFeedback. This advantage is evident on alignment benchmarks such as AlpacaEval, ArenaHard, and WildBench.
</details><be>

## Dataset Details
This dataset is generated by [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) using [Magpie](https://huggingface.co/Magpie-Align). Please refer to our [paper](https://arxiv.org/abs/2406.08464) and [codebase](https://github.com/magpie-align/magpie) for implementation details.

This is the filtered data. Please see below for the filter design. Please do not use **Magpie-Pro-300K-Filtered** and **Magpie-Pro-MT-300K** to fine-tune the model simultaneously as they are largely the same for the first turn!

You can find the model fine-tuned using this dataset [here](https://huggingface.co/Magpie-Align/Llama-3-8B-Magpie-Pro-SFT-v0.1).

## Filter Setups
- **Input Quality**: >= average
- **Instruction Reward**: >=-10
- Remove repetition and incomplete instructions (e.g., end with :)
- Choose 300K data with the longest responses

## Dataset Navigation 🧭
|Model Name | Dataset | Type | Description |
|-------------|:-------|:-------|:-------|
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-1M](https://huggingface.co/datasets/Magpie-Align/Llama-3-Magpie-Pro-1M-v0.1) | SFT | 1M Raw conversations built with Meta Llama 3 70B.
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered) | SFT | Apply a filter and select 300K high quality conversations.
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-MT-300K](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-MT-300K-v0.1) | SFT | Select 300K difficult questions and extend to multi-turn conversations.
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-3M](https://huggingface.co/datasets/Magpie-Align/Llama-3-Magpie-Air-3M-v0.1) | SFT | 3M Raw conversations built with Meta Llama 3 8B.
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Air-300K-Filtered) | SFT | Apply a filter and select 300K high quality data.
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-MT-300K](https://huggingface.co/datasets/Magpie-Align/Magpie-Air-MT-300K-v0.1) | SFT | Select 300K difficult questions and extend to multi-turn conversations.


项目官网：[https://magpie-align.github.io/](https://magpie-align.github.io/)

arXiv技术报告：[https://arxiv.org/abs/2406.08464](https://arxiv.org/abs/2406.08464)

代码仓库：[https://github.com/magpie-align/magpie](https://github.com/magpie-align/magpie)

## 摘要
<details><summary>点击展开</summary>
高质量指令数据对于对齐大语言模型（Large Language Model，LLM）至关重要。尽管部分模型（如Llama-3-Instruct）已开源权重，但其对齐数据仍处于私有状态，这阻碍了人工智能的民主化进程。现有开源数据构建方法面临人工成本高昂、提示预设范围有限的问题，难以有效扩展，可能限制了公共对齐数据集的多样性与质量。我们能否直接从已对齐的大语言模型中提取信息，从而大规模合成高质量指令数据？为此，我们提出了一种用于生成大规模对齐数据的自合成方法，命名为喜鹊（Magpie）。我们的核心观察是：得益于自回归特性，当仅向Llama-3-Instruct等已对齐大语言模型输入用户消息预留位置之前的左侧模板时，模型可生成用户查询。我们利用该方法对Llama-3-Instruct进行提示，生成了400万条指令及其对应的响应。我们对提取的数据进行了全面分析，并筛选出30万个高质量实例。为对比Magpie数据集与其他公共指令数据集的效果，我们分别使用各数据集对Llama-3-8B-Base进行微调，并评估微调后模型的性能。结果表明，在部分任务中，使用Magpie数据微调的模型性能可与官方Llama-3-8B-Instruct相媲美——尽管后者通过监督微调（Supervised Fine-Tuning, SFT）与后续反馈学习，使用了1000万条数据进行增强。我们还证实，仅使用Magpie数据进行监督微调，其性能可超越此前同时用于监督微调与偏好优化的公共数据集（如结合UltraFeedback的直接偏好优化）。这一优势在AlpacaEval、ArenaHard与WildBench等对齐基准测试中尤为显著。
</details><br>

## 数据集详情
本数据集由[Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct)通过[Magpie](https://huggingface.co/Magpie-Align)生成。有关实现细节，请参阅我们的[论文](https://arxiv.org/abs/2406.08464)与[代码库](https://github.com/magpie-align/magpie)。
本数据集为经过筛选的数据。有关筛选设计详见下文。请勿同时使用**Magpie-Pro-300K-Filtered**与**Magpie-Pro-MT-300K**进行模型微调，二者的第一轮对话内容高度重合。
你可在[此处](https://huggingface.co/Magpie-Align/Llama-3-8B-Magpie-Pro-SFT-v0.1)找到使用本数据集微调得到的模型。

## 筛选设置
- **输入质量**：≥平均值
- **指令奖励**：≥-10
- 移除重复与不完整的指令（例如以冒号结尾的内容）
- 选取30万个响应最长的数据

## 数据集导航 🧭
| 模型名称 | 数据集 | 类型 | 描述 |
|---------|-------|-----|-----|
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-1M](https://huggingface.co/datasets/Magpie-Align/Llama-3-Magpie-Pro-1M-v0.1) | 监督微调 | 基于Meta Llama 3 70B构建的100万条原始对话数据。
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-300K-Filtered) | 监督微调 | 经过筛选后选取的30万个高质量对话数据。
| [Llama 3 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) | [Magpie-Pro-MT-300K](https://huggingface.co/datasets/Magpie-Align/Magpie-Pro-MT-300K-v0.1) | 监督微调 | 选取30万个高难度问题并扩展至多轮对话。
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-3M](https://huggingface.co/datasets/Magpie-Align/Llama-3-Magpie-Air-3M-v0.1) | 监督微调 | 基于Meta Llama 3 8B构建的300万条原始对话数据。
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-300K-Filtered](https://huggingface.co/datasets/Magpie-Align/Magpie-Air-300K-Filtered) | 监督微调 | 经过筛选后选取的30万个高质量数据。
| [Llama 3 8B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) | [Magpie-Air-MT-300K](https://huggingface.co/datasets/Magpie-Align/Magpie-Air-MT-300K-v0.1) | 监督微调 | 选取30万个高难度问题并扩展至多轮对话。