deita-10k-v0

<img src="https://huggingface.co/datasets/hkust-nlp/deita-images/resolve/main/logo-final.png" alt="Deita banner" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>

# Dataset Card for Deita 10K V0

[GitHub](https://github.com/hkust-nlp/deita) | [Paper](https://arxiv.org/abs/2312.15685)

Deita is an open-sourced project designed to facilitate **Automatic Data Selection** for instruction tuning in Large Language Models (LLMs).

This dataset includes 10k of **lightweight, high-quality** alignment SFT data, mainly automatically selected from the following datasets:

- [ShareGPT](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered) (Apache 2.0 listed, no official repo found): Use the 58 K ShareGPT dataset for selection.
- [UltraChat](https://huggingface.co/datasets/stingning/ultrachat) (MIT): Sample 105 K UltraChat dataset for selection.
- [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k) : Use the evolved data of Alpaca and ShareGPT with 143 K mixture for selection.

**Model Family**: Other models and the dataset are found in the [Deita Collection](https://huggingface.co/collections/hkust-nlp/deita-6569c198c174808d94cf5bd4)

## Performance

| Model | Align | Data Size | MT-Bench | AlpacaEval(%) | OpenLLM (Avg.) |
|------------------------------------------------|-----------|------------|----------|---------------|----------------|
| **Proprietary Models** | | | | | |
| GPT-4-Turbo | ? | -- | 9.32 | 97.70 | -- |
| GPT-4 | SFT + PPO | -- | 8.99 | 95.03 | -- |
| Claude-2 | SFT + PPO | -- | 8.06 | 91.36 | -- |
| GPT-3.5-turbo | SFT + PPO | -- | 7.94 | 89.37 | -- |
| **Open-sourced Models based on LLaMA-1-13B** | | | | | |
| LIMA | SFT | 1K SFT | 4.29 | 41.98 | 59.82 |
| WizardLM-13B | SFT | 70K SFT | 6.35 | 75.31 | 58.96 |
| Vicuna-13B-v1.3 | SFT | 125K SFT | 6.39 | 82.11 | 60.01 |
| Random | SFT | 10K SFT | 6.03 | 71.52 | 60.14 |
| DEITA-LLaMA1-13B-v1.0-sft | SFT | 10K SFT | 6.60 | 78.01 | 64.27 |
| **Open-sourced Models based on LLaMA-2-13B** | | | | | |
| Tulu-2-13B | SFT | 326K SFT | 6.70 | 78.90 | -- |
| Tulu-2-13B+DPO | SFT + DPO | 326K SFT + 60K DPO | 7.00 | 89.50 | -- |
| LLaMA2-13B-Chat | SFT + PPO | -- | 6.65 | 81.09 | -- |
| WizardLM-13B-v1.2 | SFT | >70K SFT | 7.09 | 89.17 | -- |
| Vicuna-13B-v1.5 | SFT | 125K SFT | 6.57 | 78.80 | 61.63 |
| Random | SFT | 10K SFT | 5.78 | 65.19 | 61.32 |
| DEITA-LLaMA2-13B-v1.0-sft | SFT | 10K SFT | 6.79 | 81.09 | 62.71 |
| **Open-sourced Models based on Mistral-7B** | | | | | |
| Mistral-7B-Instruct-v0.1 | -- | -- | 6.84 | 69.65 | 60.45 |
| Zephyr-7B-sft | SFT | 200K SFT | 5.32 | 75.12 | 60.93 |
| $\text{Zephyr-7B-}\beta$ | SFT + DPO | 200K SFT + 60K DPO | 7.34 | 90.60 | 66.36 |
| OpenChat-3.5 | C-RLFT | >> 70K C-RLFT | 7.81 | 88.51 | -- |
| Starling-7B | C-RLFT + APA | >>70K C-RLFT + 183K APA | 8.09 | 91.99 | -- |
| Random | SFT | 10K SFT | 5.89 | 56.90 | 61.72 |
| DEITA-7B-v1.0-sft (6K) | SFT | 6K SFT | 7.22 | 80.78 | 64.94 |
| DEITA-7B-v1.0-sft (10K) | SFT | 10K SFT | 7.32 | 81.67 | 64.00 |
| DEITA-7B-v1.0 | SFT + DPO | 6K SFT + 10K DPO | 7.55 | 90.06 | 69.86 |

## Citation
If you find the content of this project helpful, please cite our paper as follows:

```
@misc{liu2023what,
title={What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning},
author={Wei Liu and Weihao Zeng and Keqing He and Yong Jiang and Junxian He},
year={2023},
eprint={2312.15685},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```

<img src="https://huggingface.co/datasets/hkust-nlp/deita-images/resolve/main/logo-final.png" alt="Deita banner" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>

# Deita 10K V0 数据集卡片

[GitHub](https://github.com/hkust-nlp/deita) | [论文](https://arxiv.org/abs/2312.15685)

Deita是一个开源项目，旨在助力大语言模型（Large Language Model，LLM）指令微调中的**自动数据选择**。

本数据集包含10k条**轻量、高质量**的对齐监督微调（Supervised Fine-Tuning，SFT）数据，主要从以下数据集自动筛选而来：
- [ShareGPT](https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered)（遵循Apache 2.0协议，未找到官方仓库）：选取58k条ShareGPT数据集用于筛选。
- [UltraChat](https://huggingface.co/datasets/stingning/ultrachat)（遵循MIT协议）：采样105k条UltraChat数据集用于筛选。
- [WizardLM](https://huggingface.co/datasets/WizardLM/WizardLM_evol_instruct_V2_196k)：使用由Alpaca与ShareGPT演化而来的143k混合数据集用于筛选。

**模型家族**：其他模型与数据集可在 [Deita 合集](https://huggingface.co/collections/hkust-nlp/deita-6569c198c174808d94cf5bd4) 中获取。

## 模型性能

| 模型 | 对齐方式 | 数据规模 | MT-Bench | AlpacaEval(%) | OpenLLM（平均分） |
|------------------------------------------------|-----------|------------|----------|---------------|----------------|
| **专有模型** | | | | | |
| GPT-4-Turbo | 未知 | -- | 9.32 | 97.70 | -- |
| GPT-4 | 监督微调+近端策略优化（Proximal Policy Optimization，PPO） | -- | 8.99 | 95.03 | -- |
| Claude-2 | 监督微调+近端策略优化（Proximal Policy Optimization，PPO） | -- | 8.06 | 91.36 | -- |
| GPT-3.5-turbo | 监督微调+近端策略优化（Proximal Policy Optimization，PPO） | -- | 7.94 | 89.37 | -- |
| **基于LLaMA-1-13B的开源模型** | | | | | |
| LIMA | 监督微调 | 1k条监督微调数据 | 4.29 | 41.98 | 59.82 |
| WizardLM-13B | 监督微调 | 70k条监督微调数据 | 6.35 | 75.31 | 58.96 |
| Vicuna-13B-v1.3 | 监督微调 | 125k条监督微调数据 | 6.39 | 82.11 | 60.01 |
| Random | 监督微调 | 10k条监督微调数据 | 6.03 | 71.52 | 60.14 |
| DEITA-LLaMA1-13B-v1.0-sft | 监督微调 | 10k条监督微调数据 | 6.60 | 78.01 | 64.27 |
| **基于LLaMA-2-13B的开源模型** | | | | | |
| Tulu-2-13B | 监督微调 | 326k条监督微调数据 | 6.70 | 78.90 | -- |
| Tulu-2-13B+DPO | 监督微调+直接偏好优化（Direct Preference Optimization，DPO） | 326k条监督微调数据 + 60k条DPO数据 | 7.00 | 89.50 | -- |
| LLaMA2-13B-Chat | 监督微调+近端策略优化（Proximal Policy Optimization，PPO） | -- | 6.65 | 81.09 | -- |
| WizardLM-13B-v1.2 | 监督微调 | >70k条监督微调数据 | 7.09 | 89.17 | -- |
| Vicuna-13B-v1.5 | 监督微调 | 125k条监督微调数据 | 6.57 | 78.80 | 61.63 |
| Random | 监督微调 | 10k条监督微调数据 | 5.78 | 65.19 | 61.32 |
| DEITA-LLaMA2-13B-v1.0-sft | 监督微调 | 10k条监督微调数据 | 6.79 | 81.09 | 62.71 |
| **基于Mistral-7B的开源模型** | | | | | |
| Mistral-7B-Instruct-v0.1 | 无对齐方式 | -- | 6.84 | 69.65 | 60.45 |
| Zephyr-7B-sft | 监督微调 | 200k条监督微调数据 | 5.32 | 75.12 | 60.93 |
| $ ext{Zephyr-7B-}eta$ | 监督微调+DPO | 200k条监督微调数据 + 60k条DPO数据 | 7.34 | 90.60 | 66.36 |
| OpenChat-3.5 | C-RLFT | >>70k条C-RLFT数据 | 7.81 | 88.51 | -- |
| Starling-7B | C-RLFT + APA | >>70k条C-RLFT数据 + 183k条APA数据 | 8.09 | 91.99 | -- |
| Random | 监督微调 | 10k条监督微调数据 | 5.89 | 56.90 | 61.72 |
| DEITA-7B-v1.0-sft (6K) | 监督微调 | 6k条监督微调数据 | 7.22 | 80.78 | 64.94 |
| DEITA-7B-v1.0-sft (10K) | 监督微调 | 10k条监督微调数据 | 7.32 | 81.67 | 64.00 |
| DEITA-7B-v1.0 | 监督微调+DPO | 6k条监督微调数据 + 10k条DPO数据 | 7.55 | 90.06 | 69.86 |

## 引用
若您认为本项目内容有所助益，请按以下格式引用我们的论文：

@misc{liu2023what,
title={What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning},
author={Wei Liu and Weihao Zeng and Keqing He and Yong Jiang and Junxian He},
year={2023},
eprint={2312.15685},
archivePrefix={arXiv},
primaryClass={cs.CL}
}