BigDong commited on
Commit
7ddd417
·
1 Parent(s): f5030f4

update README

Browse files
Files changed (2) hide show
  1. README.md +20 -14
  2. README_ZH.md +108 -0
README.md CHANGED
@@ -36,29 +36,35 @@ configs:
36
  <img src="assets/ultra-fineweb-logo.png" width="600"/>
37
  </div>
38
 
39
- <!-- <div align="center">
40
- English | [简体中文]()
41
- </div> -->
42
-
43
- <div align="center">
44
-
45
- [📜 Ultra-FineWeb Technical Report](https://arxiv.org/abs/2505.05427) | [📄 MiniCPM4 Paper](https://huggingface.co/papers/2506.07900) | [💻 GitHub Repository](https://github.com/openbmb/minicpm) | [🌐 MiniCPM4 Project Page](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b)
46
-
47
- </div>
 
 
 
48
 
49
  ## 📚 Introduction
50
 
51
- Ultra-FineWeb is a **large-scale, high-quality, and efficiently-filtered dataset**. We use the proposed efficient verification-based high-quality filtering pipeline to the FineWeb and Chinese FineWeb datasets (source data from Chinese FineWeb-edu-v2, which includes IndustryCorpus2, MiChao, WuDao, SkyPile, WanJuan, ChineseWebText, TeleChat, and CCI3), resulting in the creation of higher-quality Ultra-FineWeb-en with approximately 1T tokens, and Ultra-FineWeb-zh datasets with approximately 120B tokens, collectively referred to as Ultra-FineWeb. ***Ultra-FineWeb*** serves as a core pre-training web dataset for the [MiniCPM4 Series](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) models.
52
  - [Ultra-FineWeb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb): Ultra-FineWeb, a **large-scale, high-quality, and efficiently-filtered dataset**, with 1T English tokens and 120B Chinese tokens. (**<-- you are here**)
53
  - [Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier): Ultra-FineWeb classifier, for filtering high-quality data from web corpora.
 
54
 
55
  ## 📢 What's New
56
 
57
- - **[2025.05.09]** **Ultra-FineWeb** technical report is available on [arXiv](https://arxiv.org/abs/2505.05427). 🔥🔥🔥
58
- - **[2025.05.15]** **Ultra-FineWeb** tops the Hugging Face Datasets Trending list, reaching the #1 spot! ⭐️⭐️⭐️
59
- - **[2025.06.06]** **Ultra-FineWeb-en** and **Ultra-FineWeb-zh** datasets are now available on Hugging Face, released alongside the [MiniCPM4 Series](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) models.
60
  - **[2025.06.16]** The **Ultra-FineWeb-classifier** is now available on Hugging Face: [openbmb/Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier).
61
- - **[2025.12.10]** [***Ultra-FineWeb-en-v1.4***](https://huggingface.co/datasets/openbmb/Ultra-FineWeb/tree/main/data/ultrafineweb_en_v1_4) is released! 2.2T tokens fully open-sourced! Built on [FineWeb-v1.4](https://huggingface.co/datasets/HuggingFaceFW/fineweb), incorporating CommonCrawl snapshots from Apr 2024 - Jun 2025 to capture the latest world knowledge. 🚀🚀🚀
 
 
62
 
63
  ## 💡 Highlights
64
 
 
36
  <img src="assets/ultra-fineweb-logo.png" width="600"/>
37
  </div>
38
 
39
+ <p align="center">
40
+ <a href="https://arxiv.org/abs/2505.05427">📜 Technical Report</a> |
41
+ <a href="https://huggingface.co/collections/openbmb/ultradata">📦 UltraData Collection</a> |
42
+ <a href="https://ultradata.openbmb.cn/">🌐 UltraData</a> |
43
+ <a href="https://huggingface.co/collections/openbmb/minicpm4">🤗 MiniCPM4 Series</a> |
44
+ <a href="https://huggingface.co/collections/openbmb/minicpm5">🤗 MiniCPM5 Series</a>
45
+ </p>
46
+
47
+ <p align="center">
48
+ English |
49
+ <a href="https://huggingface.co/datasets/openbmb/Ultra-FineWeb/blob/main/README_ZH.md">中文</a>
50
+ </p>
51
 
52
  ## 📚 Introduction
53
 
54
+ Ultra-FineWeb is a **large-scale, high-quality, and efficiently-filtered dataset**. We use the proposed efficient verification-based high-quality filtering pipeline to the FineWeb and Chinese FineWeb datasets (source data from Chinese FineWeb-edu-v2, which includes IndustryCorpus2, MiChao, WuDao, SkyPile, WanJuan, ChineseWebText, TeleChat, and CCI3), resulting in the creation of higher-quality Ultra-FineWeb-en with approximately 1T tokens, and Ultra-FineWeb-zh datasets with approximately 120B tokens, collectively referred to as Ultra-FineWeb. ***Ultra-FineWeb*** serves as a core pre-training web dataset for the [MiniCPM4 Series](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) and [MiniCPM5 Series](https://huggingface.co/collections/openbmb/minicpm5) models.
55
  - [Ultra-FineWeb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb): Ultra-FineWeb, a **large-scale, high-quality, and efficiently-filtered dataset**, with 1T English tokens and 120B Chinese tokens. (**<-- you are here**)
56
  - [Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier): Ultra-FineWeb classifier, for filtering high-quality data from web corpora.
57
+ - [Ultra-FineWeb-L3](https://huggingface.co/datasets/openbmb/Ultra-FineWeb-L3): the **L3 refined data** based on **Ultra-FineWeb**, via **Q&A Pair Generation** and **Multi-style Rewriting**, with **400B+ English** and **200B+ Chinese** tokens—to our best knowledge, the largest open-source Chinese pre-training synthetic corpus to date.
58
 
59
  ## 📢 What's New
60
 
61
+ - **[2026.05.28]** The [***Ultra-FineWeb-L3***](https://huggingface.co/datasets/openbmb/Ultra-FineWeb-L3) dataset is released! The **L3 refined data** of **Ultra-FineWeb** via **Q&A Pair Generation** and **Multi-style Rewriting**, with **400B+ English** and **200B+ Chinese** tokens. To our best knowledge, it is the largest open-source Chinese pre-training synthetic corpus to date. 🚀🚀🚀
62
+ - **[2026.05.25]** ***[MiniCPM5-1B](https://huggingface.co/openbmb/MiniCPM5-1B) is released!***, the first model in the MiniCPM5 series. It is a dense 1B Transformer built for on-device, local deployment, and resource-constrained scenarios, reaching 1B-class open-source SOTA. Ultra-FineWeb serves as the core pre-training web dataset for MiniCPM5-1B.
63
+ - **[2026.02.08]** The [***UltraData***](https://ultradata.openbmb.cn/) platform is now live, introducing the [L0-L4 tiered data management framework](https://arxiv.org/abs/2602.09003). ***Ultra-FineWeb*** serves as the **L2 selected layer** for general web data in this framework. 🔍🔍🔍
64
  - **[2025.06.16]** The **Ultra-FineWeb-classifier** is now available on Hugging Face: [openbmb/Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier).
65
+ - **[2025.06.06]** **Ultra-FineWeb-en** and **Ultra-FineWeb-zh** datasets are now available on Hugging Face, released alongside the [MiniCPM4 Series](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) models.
66
+ - **[2025.05.15]** **Ultra-FineWeb** tops the Hugging Face Datasets Trending list, reaching the #1 spot! ⭐️⭐️⭐️
67
+ - **[2025.05.09]** **Ultra-FineWeb** technical report is available on [arXiv](https://arxiv.org/abs/2505.05427). 🔥🔥🔥
68
 
69
  ## 💡 Highlights
70
 
README_ZH.md ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Ultra-FineWeb
2
+
3
+ <div align="center">
4
+ <img src="assets/ultra-fineweb-logo.png" width="600"/>
5
+ </div>
6
+
7
+ <p align="center">
8
+ <a href="https://arxiv.org/abs/2505.05427">📜 技术报告</a> |
9
+ <a href="https://huggingface.co/collections/openbmb/ultradata">📦 UltraData 合集</a> |
10
+ <a href="https://ultradata.openbmb.cn/">🌐 UltraData</a> |
11
+ <a href="https://huggingface.co/collections/openbmb/minicpm4">🤗 MiniCPM4 系列</a> |
12
+ <a href="https://huggingface.co/collections/openbmb/minicpm5">🤗 MiniCPM5 系列</a>
13
+ </p>
14
+
15
+ <p align="center">
16
+ <a href="https://huggingface.co/datasets/openbmb/Ultra-FineWeb/blob/main/README.md">English</a> |
17
+ 中文
18
+ </p>
19
+
20
+ ## 📚 简介
21
+
22
+ Ultra-FineWeb 是一个**大规模、高质量、高效过滤**的网页预训练数据集。我们将提出的高效验证式高质量过滤流水线应用于 FineWeb 与 Chinese FineWeb 数据集(中文部分源数据来自 Chinese FineWeb-edu-v2,涵盖 IndustryCorpus2、MiChao、WuDao、SkyPile、WanJuan、ChineseWebText、TeleChat 与 CCI3 等),得到更高质量的 Ultra-FineWeb-en(约 1T tokens)与 Ultra-FineWeb-zh(约 120B tokens),合称 ***Ultra-FineWeb***。***Ultra-FineWeb*** 是 [MiniCPM4 系列](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) 与 [MiniCPM5 系列](https://huggingface.co/collections/openbmb/minicpm5) 的核心预训练网页数据集。
23
+
24
+ - [Ultra-FineWeb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb):大规模、高质量、高效过滤的网页预训练数据集,包含 1T 英文 tokens 与 120B 中文 tokens。(**当前数据集**)
25
+ - [Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier):Ultra-FineWeb 分类器,用于从网页语料中筛选高质量数据。
26
+ - [Ultra-FineWeb-L3](https://huggingface.co/datasets/openbmb/Ultra-FineWeb-L3):基于 Ultra-FineWeb 的 **L3 合成与增强数据**,通过**问答对生成**与**多风格改写**构建语料,包含 **400B+ 英文**与 **200B+ 中文** tokens;据我们所知,这是目前规模最大的开源中文预训练合成语料。
27
+
28
+ ## 📢 最新动态
29
+
30
+ - **[2026.05.28]** [***Ultra-FineWeb-L3***](https://huggingface.co/datasets/openbmb/Ultra-FineWeb-L3) 数据集正式发布!作为 **Ultra-FineWeb** 的 **L3 合成与增强数据**,通过**问答对生成**与**多风格改写**构建语料,包含 **400B+ 英文**与 **200B+ 中文** tokens。据我们所知,这是目前规模最大的开源中文预训练合成语料。🚀🚀🚀
31
+ - **[2026.05.25]** ***[MiniCPM5-1B](https://huggingface.co/openbmb/MiniCPM5-1B) 正式发布!*** 作为 MiniCPM5 系列的首款模型,这是一款面向端侧、本地部署与资源受限场景的 1B 稠密 Transformer,在 1B 级开源模型中达到 SOTA。Ultra-FineWeb 是 MiniCPM5-1B 的核心预训练网页数据集。
32
+ - **[2026.02.08]** [***UltraData***](https://ultradata.openbmb.cn/) 平台正式上线,并发布 [L0-L4 分级数据管理框架](https://arxiv.org/abs/2602.09003)。***Ultra-FineWeb*** 是该框架下通用网页数据的 **L2 精选层**实践。🔍🔍🔍
33
+ - **[2025.06.16]** **Ultra-FineWeb-classifier** 已在 Hugging Face 上线:[openbmb/Ultra-FineWeb-classifier](https://huggingface.co/openbmb/Ultra-FineWeb-classifier)。
34
+ - **[2025.06.06]** **Ultra-FineWeb-en** 与 **Ultra-FineWeb-zh** 数据集已在 Hugging Face 发布,并与 [MiniCPM4 系列](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b) 模型同步开源。
35
+ - **[2025.05.15]** **Ultra-FineWeb** 登上 Hugging Face Datasets Trending 榜首!⭐️⭐️⭐️
36
+ - **[2025.05.09]** **Ultra-FineWeb** 技术报告已在 [arXiv](https://arxiv.org/abs/2505.05427) 发布。🔥🔥🔥
37
+
38
+ ## 💡 亮点
39
+
40
+ > **摘要:** 随着大语言模型(LLM)的快速发展,数据质量已成为提升模型性能的关键因素。模型驱动的数据过滤逐渐成为获取高质量数据的主要途径,但仍面临两大挑战:(1)缺乏高效的数据验证策略,难以及时反馈数据质量;(2)训练分类器所用的种子数据选择缺乏明确标准,过度依赖人工经验,带来主观性。针对第一个挑战,我们提出高效验证策略,以极低计算成本快速评估数据对 LLM 训练的影响。针对第二个挑战,我们在「高质量种子数据有利于 LLM 训练」的假设下,结合上述验证策略优化正负样本选择,并提出高效数据过滤流水线。该流水线不仅提升过滤效率、分类器质量与鲁棒性,还显著降低实验与推理成本。此外,我们采用基于 *fastText* 的轻量分类器,将过滤流水线成功应用于 *FineWeb* 与 *Chinese FineWeb* 两大广泛使用的预训练语料,构建更高质量的 ***Ultra-FineWeb*** 数据集。***Ultra-FineWeb*** 包含约 1 万亿(T)英文 tokens 与 1200 亿(B)中文 tokens。实验结果表明,基于 Ultra-FineWeb 训练的 LLM 在多项基准上显著优于对照组,验证了我们流水线在提升数据质量与训练效��方面的有效性。
41
+
42
+ <div align="center">
43
+ <img src="assets/ultra-fineweb-pipeline.png" width="600"/>
44
+ </div>
45
+
46
+ - **高效验证策略:** 提出计算高效的数据验证策略,以极低计算成本快速评估数据对 LLM 训练性能的影响,显著提升高质量数据过滤实验的效率。
47
+ - **大规模高质量预训练数据集:** 设计并实现高效高质量数据过滤流水线,应用于 FineWeb 与 Chinese FineWeb,构建更高质量的数据集,助力高质量 LLM 训练。
48
+ - **轻量分类器:** Ultra-FineWeb 分类器显著降低推理成本,在同一数据源提取文本上表现更优,验证了所提数据过滤流水线在提升数据质量与训练效率方面的有效性。
49
+
50
+ ## 📈 评测结果
51
+
52
+ 我们采用 MiniCPM-1.2B 模型架构与 MiniCPM3-4B 词表,每个实验训练 100B tokens,在计算可控的前提下完成全面的数据性能验证。评测使用 Lighteval 库,在 11 个基准上评估训练后的模型,所有指标均为 zero-shot 设置,包括:
53
+
54
+ - **英文基准:** MMLU、ARC-C、ARC-E、CommonSenseQA、HellaSwag、OpenbookQA、PIQA、SIQA、Winogrande。
55
+ - **中文基准:** C-Eval、CMMLU。
56
+
57
+ 详细评测结果如下:
58
+
59
+ - **单数据集实验。** 对各数据集分别独立训练,便于在相同来源下直接对比不同处理方案的数据效果。
60
+ <img src="assets/individual-english-table.png" alt="英文单数据集结果表" width="75%">
61
+ <img src="assets/individual-chinese-table.png" alt="中文单数据集结果表" width="75%">
62
+ <img src="assets/individual-plot.png" alt="单数据集实验曲线" width="100%">
63
+
64
+ - **混合数据实验。** 采用 60% 英文数据、30% 中文数据与 10% 代码数据(StarCoder-v2)的混合配置。
65
+ <img src="assets/mix-table.png" alt="混合数据结果表" width="75%">
66
+ <img src="assets/mix-plot.png" alt="混合数据实验曲线" width="100%">
67
+
68
+ - **Loss 与性能估计结果。** 采用 [Densing Law](https://arxiv.org/abs/2412.04315) 提出的性能估计方法,进一步分析与验证 Ultra-FineWeb 的有效性。
69
+
70
+ <img src="assets/densing-law-table.png" alt="Densing Law 结果表" width="75%">
71
+ <img src="assets/densing-law-plot.png" alt="Densing Law 曲线" width="100%">
72
+
73
+ ## ❤️ 致谢
74
+
75
+ - ***Ultra-FineWeb 分类器*** 基于 [fastText](https://fasttext.cc/) 构建。
76
+ - ***Ultra-FineWeb-en 数据集*** 基于 [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) 构建。
77
+ - ***Ultra-FineWeb-zh 数据集*** 基于 [IndustryCorpus2](https://huggingface.co/datasets/BAAI/IndustryCorpus2)、[MiChao](https://opendatalab.com/OpenDataLab/MiChao)、[WuDao](https://data.baai.ac.cn/details/WuDaoCorporaText)、[SkyPile](https://huggingface.co/datasets/Skywork/SkyPile-150B)、[WanJuan](https://opendatalab.com/OpenDataLab/WanJuanCC)、[ChineseWebText](https://huggingface.co/datasets/CASIA-LM/ChineseWebText2.0)、[TeleChat](https://huggingface.co/datasets/Tele-AI/TeleChat-PTD) 与 [CCI3](https://huggingface.co/datasets/BAAI/CCI3-Data) 等数据集构建。
78
+
79
+ 感谢这些优秀开源工作,正是开源社区的贡献让 Ultra-FineWeb 成为可能!🙌
80
+
81
+ ## 🌟 引用
82
+
83
+ 若本工作对您的研究有帮助,欢迎引用:
84
+
85
+ ```bibtex
86
+ @misc{wang2025ultrafineweb,
87
+ title={{Ultra-FineWeb}: Efficient Data Filtering and Verification for High-Quality LLM Training Data},
88
+ author={Yudong Wang and Zixuan Fu and Jie Cai and Peijun Tang and Hongya Lyu and Yewei Fang and Zhi Zheng and Jie Zhou and Guoyang Zeng and Chaojun Xiao and Xu Han and Zhiyuan Liu},
89
+ year={2025},
90
+ eprint={2505.05427},
91
+ archivePrefix={arXiv},
92
+ primaryClass={cs.CL},
93
+ }
94
+ ```
95
+
96
+ 以及使用 Ultra-FineWeb 的主要论文:
97
+
98
+ ```bibtex
99
+ @article{minicpm4,
100
+ title={MiniCPM4: Ultra-Efficient LLMs on End Devices},
101
+ author={MiniCPM Team},
102
+ year={2025}
103
+ }
104
+ ```
105
+
106
+ ## 💳 许可证
107
+
108
+ 本项目基于 [Apache 2.0](./LICENSE) 许可证发布。请注意,***Ultra-FineWeb*** 由多个数据集构建而成,使用者应**分别查阅各源数据集的 LICENSE**,以确保合规使用。