anas-awadalla commited on
Commit
2e16c4c
β€’
1 Parent(s): 968a328

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +149 -0
README.md ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - image-to-text
5
+ - text-generation
6
+ language:
7
+ - en
8
+ tags:
9
+ - multimodal
10
+ pretty_name: MINT-1T
11
+ size_categories:
12
+ - 100B<n<1T
13
+ ---
14
+
15
+ <h1 align="center">
16
+ πŸƒ MINT-1T:<br>Scaling Open-Source Multimodal Data by 10x:<br> A Multimodal Dataset with One Trillion Tokens
17
+ </h1>
18
+
19
+ πŸƒ MINT-1T is an open-source **M**ultimodal **INT**erleaved dataset with 1 trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. πŸƒ MINT-1T is designed to facilitate research in multimodal pretraining. πŸƒ MINT-1T is created by a team from the University of Washington in collaboration with Salesforce Research, other academic institutions including Stanford University, University of Texas at Austin, and University of California Berkeley.
20
+
21
+ You are currently viewing a subset of the PDF portion of πŸƒ MINT-1T associated with CommonCrawl dump `CC-2023-23`. For other PDF, HTML, and ArXiv subsets, refer to the [πŸƒ MINT-1T collection](https://huggingface.co/collections/mlfoundations/mint-1t-6690216ca4d0df7e518dde1c).
22
+
23
+ ## Dataset Details
24
+
25
+ ### Dataset Sources
26
+
27
+ - **Repository**: https://github.com/mlfoundations/MINT-1T
28
+ - **Paper:** https://arxiv.org/abs/2406.11271
29
+ - **Blog:**
30
+
31
+ ## Uses
32
+
33
+ ### Direct Use
34
+
35
+ <!-- This section describes suitable use cases for the dataset. -->
36
+
37
+ πŸƒ MINT-1T is designed to facilitate research in multimodal pretraining. The dataset can be used for training multimodal models that can reson about interleaved text and images sequences such as [Idefics2](https://huggingface.co/HuggingFaceM4/idefics2-8b), [XGen-MM](https://huggingface.co/Salesforce/xgen-mm-phi3-mini-instruct-r-v1), and [Chameleon](https://huggingface.co/facebook/chameleon-30b).
38
+
39
+ ### Out-of-Scope Use
40
+
41
+ <!-- This section addresses misuse, malicious use, and uses that the dataset will not work well for. -->
42
+
43
+ πŸƒ MINT-1T was built to make research into large multimodal models more accessible. Using
44
+ the dataset to train models that ingest or generate personally identifying information (such
45
+ as images of people’s faces and other sensitive content) as well as military applications are all inappropriate use cases of πŸƒ MINT-1T.
46
+
47
+ ## Dataset Creation
48
+
49
+ ### Curation Rationale
50
+
51
+ πŸƒ MINT-1T was created to address a significant gap in the open-source domain by providing a large-scale multimodal interleaved dataset for pre-training large multimodal models. This dataset aims to be a valuable resource for the research community, facilitating open science in multimodal pretraining.
52
+
53
+ ### Source Data
54
+
55
+ The dataset is a comprehensive collection of multimodal documents from various sources:
56
+
57
+ - HTML documents: Filtered from CommonCrawl WARC dumps spanning from 2017 to 2024
58
+ - PDF documents: Extracted from CommonCrawl WAT dumps covering 2023 to 2024
59
+ - ArXiv documents: A subset of papers from the ArXiv repository
60
+
61
+ In total, πŸƒ MINT-1T contains 1056.8 million documents, broken down as follows:
62
+ - 1029.4 million HTML documents
63
+ - 26.8 million PDF documents
64
+ - 0.6 million ArXiv documents
65
+
66
+ #### Data Collection and Processing
67
+
68
+ The data collection and processing involved several steps:
69
+
70
+ 1. Document Extraction:
71
+ - HTML documents were parsed from CommonCrawl WARC files
72
+ - PDF documents were extracted from CommonCrawl WAT files
73
+ - ArXiv papers were directly sourced from ArXiv S3 buckets
74
+
75
+ 2. Filtering Process:
76
+ - Applied text quality filters to ensure content relevance and readability
77
+ - Removed duplicate content at both paragraph and document levels
78
+ - Filtered out undesirable content based on predefined criteria
79
+ - Verified image availability and quality for HTML documents
80
+ - Limited PDF size to 50MB and 50 pages to manage dataset size and quality
81
+
82
+ 3. Image Processing:
83
+ - Used NSFW image detection to remove pornographic or otherwise undesirable images
84
+ - Removed images smaller than 150 pixels or larger than 20,000 pixels
85
+ - Adjusted aspect ratio thresholds for HTML (2:1) and PDF (3:1) to preserve scientific figures
86
+
87
+ 4. Text Processing:
88
+ - Used fasttext for language identification, focusing on English content
89
+ - Masked personally identifiable information such as email addresses and IP addresses
90
+ - Applied paragraph and document-level deduplication using Bloom filters
91
+
92
+ 5. PDF Specific Processing:
93
+ - Used PyMuPDF for parsing PDFs and extracting reading order
94
+ - Clustered text blocks based on columns and ordered from top left to bottom right
95
+
96
+ 6. ArXiv Specific Processing:
97
+ - Used TexSoup to parse LaTeX source code and interleave images with text
98
+ - Cleaned up LaTeX code by removing imports, bibliography, tables, and citation tags
99
+
100
+ Various open-source tools were utilized in this process, including fasttext, [PyMuPDF](https://github.com/pymupdf/PyMuPDF), and [DCLM](https://www.datacomp.ai/dclm/) and [bff](https://github.com/revbucket/bff) for deduplication and content filtering.
101
+
102
+ #### Personal and Sensitive Information
103
+
104
+ Despite sourcing from public web data, significant efforts were made to minimize the inclusion of personal and sensitive information:
105
+
106
+ - Email addresses and IP addresses were masked to protect privacy
107
+ - An NSFW image classifierto remove inappropriate visual content
108
+ - URLs containing substrings associated with undesirable or sensitive content were filtered out
109
+
110
+ However, users should be aware that as the data originates from the public web, it may still contain some sensitive or personal information. The dataset creators acknowledge this limitation and advise users to exercise caution and potentially apply additional filtering based on their specific use cases.
111
+
112
+ ## Bias, Risks, and Limitations
113
+
114
+ Several potential biases, risks, and limitations have been identified:
115
+
116
+ 1. Data Bias: As the dataset is sourced from web crawls, it may inherit biases present in online content. This could include demographic, cultural, or topical biases.
117
+
118
+ 2. Content Risks: Despite extensive filtering, there's a possibility that some offensive, insensitive, or inappropriate content may remain in the dataset.
119
+
120
+ 3. Image Availability: The dataset relies on external image URLs, which may become unavailable over time due to link rot, potentially affecting the dataset's long-term usability.
121
+
122
+ 4. PDF Parsing Limitations: The current method for extracting reading order from PDFs may not always accurately capture the intended flow, especially for documents with complex layouts.
123
+
124
+ 5. Potential Legal and Ethical Concerns: While efforts were made to respect robots.txt files and remove sensitive information, there may still be content that individuals did not explicitly consent to include.
125
+
126
+ ### Recommendations
127
+
128
+ Given these considerations, the following recommendations are provided:
129
+
130
+ 1. Additional Filtering: Users are strongly encouraged to apply additional filtering based on their specific use case and ethical considerations.
131
+
132
+ 2. Inappropriate Use Cases: The dataset is not recommended for applications involving the processing or generation of personally identifying information, nor for military applications.
133
+
134
+ 3. Legal Compliance: Users should independently verify compliance with applicable laws before employing MINT-1T for commercial purposes.
135
+
136
+ 4. Bias Awareness: Researchers and developers should be cognizant of potential biases in the dataset and consider their impact on model training and outputs.
137
+
138
+ ## License
139
+ We release πŸƒ MINT-1T under a CC-BY-4.0 license, designating it primarily as a research artifact. While the dataset is freely available, users are responsible for ensuring its legal use in commercial settings. Users must independently verify compliance with applicable laws before employing MINT-1T for commercial purposes.
140
+
141
+ ## Citation
142
+
143
+ ```
144
+ @article{awadalla2024mint1t,
145
+ title={MINT-1T: Scaling Open-Source Multimodal Data by 10x: A Multimodal Dataset with One Trillion Tokens},
146
+ author={Anas Awadalla and Le Xue and Oscar Lo and Manli Shu and Hannah Lee and Etash Kumar Guha and Matt Jordan and Sheng Shen and Mohamed Awadalla and Silvio Savarese and Caiming Xiong and Ran Xu and Yejin Choi and Ludwig Schmidt},
147
+ year={2024}
148
+ }
149
+ ```