TianyiQ commited on
Commit
08e419d
1 Parent(s): 04e63de

Upload new README

Browse files
README.md ADDED
@@ -0,0 +1,175 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ - de
8
+ - fr
9
+ size_categories:
10
+ - 1M<n<10M
11
+ ---
12
+
13
+ # ProgressGym-HistText
14
+
15
+ ## Overview
16
+
17
+ #### The ProrgressGym Framework
18
+
19
+ ![Framework Diagram](./readme-assets/main-diagram.png)
20
+
21
+ **ProgressGym-HistText** is part of the **ProgressGym** framework for research and experimentation on *progress alignment* - the emulation of moral progress in AI alignment algorithms, as a measure to prevent risks of societal value lock-in.
22
+
23
+ To quite the paper *ProgressGym: Alignment with a Millennium of Moral Progress*:
24
+
25
+ > Frontier AI systems, including large language models (LLMs), hold increasing influence over the epistemology of human users. Such influence can reinforce prevailing societal values, potentially contributing to the lock-in of misguided moral beliefs and, consequently, the perpetuation of problematic moral practices on a broad scale.
26
+ >
27
+ > We introduce *progress alignment* as a technical solution to mitigate this imminent risk. Progress alignment algorithms learn to emulate the mechanics of human moral progress, thereby addressing the susceptibility of existing alignment methods to contemporary moral blindspots.
28
+
29
+ #### The ProgressGym-HisText Dataset
30
+
31
+ ProgressGym-HistText is the central dataset in the ProgressGym framework. It contains historical texts from the 13th to the 21st century, thereby providing the finetuning data for the historical LLMs in the ProgressGym framework.
32
+
33
+ To mitigate the problems of mislabeling, OCR errors, and other quality issues in raw historical texts, ProgressGym-HistText has undergone multiple rounds of filtering and refinement, through both rule-based and machine learning-based pipelines.
34
+
35
+ We collected historical texts from portions of Internet Archive - Library of Congress, Project Gutenberg, Early English Books Online (EEBO), and [Pile of Law](https://huggingface.co/datasets/pile-of-law/pile-of-law), which are public-domain, freely available digital libraries of texts. The dataset encompasses different types of texts: fiction, nonfiction, legal, administrative, religious, and more.
36
+
37
+ The texts are organized into `.json` files by year, with each year's file containing a list of passages. Each passage is represented as a dictionary, with the fields `creation_year`, `source_dataset`, and `content` being mandatory, and many other metadata fields being optional to include.
38
+
39
+ ## Statistics
40
+
41
+ The ProgressGym-HistText dataset is 38GB in size. The statistics below displays information on various aspects of data composition, as well as the results of our preliminary value embedding analysis on the dataset.
42
+
43
+ Please note that dimensions of the value embeddings are only chosen for demonstrative purposes (without cherry-picking); the ProgressGym framework is *not* strongly anchored on these dimensions, and instead examines a wider spectrum of 19 value dimensions. Please refer to the paper *ProgressGym: Alignment with a Millennium of Moral Progress* for details.
44
+
45
+ ![image-20240630103046223](./readme-assets/data-sources.png)
46
+
47
+ ![image-20240630101921110](./readme-assets/data-stats.png)
48
+
49
+ ## Links
50
+
51
+ - **[Paper Preprint]** ProgressGym: Alignment with a Millennium of Moral Progress *(link coming soon)*
52
+ - **[Github Codebase]** PKU-Alignment/ProgressGym *(link coming soon)*
53
+ - **[Huggingface Data & Model Collection]** [PKU-Alignment/ProgressGym](https://huggingface.co/collections/PKU-Alignment/progressgym-666735fcf3e4efa276226eaa)
54
+ - **[PyPI Package]** *(coming soon)*
55
+
56
+ ## Citation
57
+
58
+ If the datasets, models, or framework of ProgressGym help you in your project, please cite ProgressGym using the bibtex entry below.
59
+
60
+ ```text
61
+ @article{progressgym,
62
+ title={ProgressGym: Alignment with a Millennium of Moral Progress},
63
+ author={Tianyi Qiu and Yang Zhang and Xuchuan Huang and Jasmine Xinze Li and Jiaming Ji and Yaodong Yang},
64
+ journal={arXiv preprint arXiv:2406.XXXXX},
65
+ eprint={2406.XXXXX},
66
+ eprinttype = {arXiv},
67
+ year={2024}
68
+ }
69
+ ```
70
+
71
+ ## Data Samples
72
+
73
+ ```json
74
+ [{
75
+ "creation_year": 1804,
76
+ "source_dataset": "Pile_of_Law",
77
+ "content": "Title: To Thomas Jefferson from Daniel Baldwin, 8 October 1804\nFrom: Baldwin, Daniel\nTo: Jefferson, Thomas\n\n\n \n\n \n Much esteemed President of the united States\n \n Paterson New jersey Oct. 8. 1804\n \n It is not for me as an individual to point out to your duty, or to direct you in any part of it. But from the probable events that I think is likely to take place Europe...",
78
+ "culture": "English",
79
+ "source_dataset_detailed": "Pile_of_Law_founding_docs",
80
+ "source_dataset_detailed_explanation": "Letters from U.S. founders.",
81
+ "url": "https://founders.archives.gov/API/docdata/Jefferson/01-44-02-0447",
82
+ "created_timestamp": "10-08-1804",
83
+ "downloaded_timestamp": "10-19-2021"
84
+ },
85
+ {
86
+ "creation_year": 1491,
87
+ "source_dataset": "EEBO",
88
+ "content": "When one of the lucky ones shall die, it is most necessary to have a special friend, who will heartily help and pray for him, and in addition, counsel the sick and ensure that all others do the same. Remember the great benefits of God bestowed upon him up to that time, and particularly the Passion of our Lord. Read some story of saints or the seven psalms with the litany or our Lady's Psalter in part or whole...",
89
+ "creation_year_earliest": 1491,
90
+ "creation_year_latest": 1491,
91
+ "source_dataset_detailed": "EEBO_Phase2"
92
+ },
93
+ {
94
+ "creation_year": 1269,
95
+ "source_dataset": "gutenberg",
96
+ "content": "TRANSCRIBER\u2019S NOTE:\nItalic text has been marked with _underscores_. [truncated] THE LADY POVERTY.\n II. How the Blessed Francis made diligent\n search for the Lady Poverty 8\n III. How two old men showed the Blessed\n Francis where he might find the\n IV. Of the First Companions of the Blessed\n V. How the Blessed Francis and his\n Companions found the Lady Poverty\n VI. The Blessed Francis and his Companions,\n exalting her virtues in\n divers ways, beseech the Lady\n Poverty to abide with them forever 28\n VII. The Answer of My Lady Poverty 41\n IX. Of the Successors of the Apostles 59\n X. That Times of Peace are unpropitious\n XII. Of the followers of a spurious Poverty...",
97
+ "source_document": "",
98
+ "culture": " English\n",
99
+ "source_dataset_detailed": "gutenberg - The Lady Poverty\n"
100
+ },
101
+ {
102
+ "creation_year": 1831,
103
+ "source_dataset": "Internet_Archive",
104
+ "content": "[An Abridgment of Elements of Criticism. By the Honorable Henry Home of Kames. Edited by John Frost, A.M.]\n\nEastern District of Pennsylvania, October 22, 1830.\n\nRemembered, that Towar & J. & D. M. Hogan have deposited in this office the title of a Book, the right whereof they claim as proprietors: \"An Abridgment of Elements of Criticism. By the Honorable Henry Home of Kames. Edited by John Frost, A.M.\"\n\nIn conformity to the act of the Congress of the United States, entitled, \"An Act for the encouragement of learning, by securing the copies of maps, charts, and books, and for establishing a public library [...] Ideas may arise in the mind without a perceived connection. We can attend to some ideas and dismiss others. Among connected objects, one suggests many of its relations; we can make a choice, electing one and rejecting others. We can insist on the slightest connection. Ideas continue through the strictest connections. The mind extends its view to a son more readily than to a servant, and to a neighbor more readily than to one living at a distance. We cannot dissolve the train, but we may vary the order. Thus, the twelve elements of criticism.\nMy wind cooling my broth,\nWould blow me to an ague, when I thought\nWhat harm a wind too great might do at sea.\nI should not see the sandy hourglass run,\nBut I should think of shallows and of flats,\nAnd see my wealthy Andrew dock'd in sand\nVailing her high top lower than her ribs,\nTo kiss her burial. Should I go to church,\nAnd see the holy edifice of stone,\nAnd not bethink me straight of dangerous rocks?\nWhich vessel touches me gently would scatter all the spices on the stream, enrobe the roaring waters with my silks, and in a word, is now worth this and now worth nothing. Merchants of Venice, Act I. Sc. 1.\n\nSome people's thoughts and circumstances crowd each other by the slightest connections. I attribute this to a bluntness in the discernment faculty; such a person has usually a great flow of ideas because they are introduced by any relations indifferently. This doctrine is vividly illustrated by Shakespeare.\n\nFalstaff: What is the gross sum that I owe you?\n\nHostess: Marry, if thou wert an honest man, thyself and thy money too. Thou didst swear to me on a gilt-parceled goblet, sitting in my Dolphin chamber, at the round table, by a sea-coal fire [...] These principles have been the declared purpose of the present undertaking.\n\nREVIEW.\n\nWhat is the general proposition to which the common proverb about taste can be resolved?\n\nHow is this proposition supported by reasoning?\n\nIs the proverb true to a certain extent?\n\nWhat is the advantage of a variety of tastes among mankind?\n\nWhat difficulties arise when applying the proverb to every subject of taste?\n\nWhat is the standard for each individual of a species?\n\nWhat conception do we form of our common nature?\n\nFor what purpose does this conviction account?\nHow is the decisive authority of this common standard illustrated? Upon what is a standard of taste erected? Is it applied to the fine arts, as well as to morals? Upon what are rules of conduct founded? Why is there not much difference of taste in the fine arts? Can a defective taste be cured? What do differences about objects of taste generally concern? What preserves uniformity of emotions and feelings among men? Do these principles always ultimately prevail?",
105
+ "title": "An abridgment of Elements of criticism",
106
+ "creator": [
107
+ "Kames, Henry Home, Lord, 1696-1782",
108
+ "Frost, John, 1800-1859, [from old catalog] ed"
109
+ ],
110
+ "subject": [
111
+ "Criticism",
112
+ "Style, Literary"
113
+ ],
114
+ "publisher": "Philadelphia, Towar, J. & D. M. Hogan; Pittsburgh, Hogan & co.",
115
+ "date": "1831",
116
+ "language": "eng",
117
+ "lccn": "11014953",
118
+ "page-progression": "lr",
119
+ "sponsor": "The Library of Congress",
120
+ "contributor": "The Library of Congress",
121
+ "scanningcenter": "capitolhill",
122
+ "mediatype": "texts",
123
+ "collection": [
124
+ "library_of_congress",
125
+ "americana"
126
+ ],
127
+ "shiptracking": "LC172",
128
+ "call_number": "9632137",
129
+ "identifier-bib": "00215413652",
130
+ "repub_state": "4",
131
+ "updatedate": "2012-11-08 22:42:22",
132
+ "updater": "ChristinaB",
133
+ "identifier": "abridgmentofelem00kame",
134
+ "uploader": "[email protected]",
135
+ "addeddate": "2012-11-08 22:42:24",
136
+ "publicdate": "2012-11-08 22:42:27",
137
+ "scanner": "scribe5.capitolhill.archive.org",
138
+ "notes": "No copyright page found.",
139
+ "repub_seconds": "2165",
140
+ "ppi": "500",
141
+ "camera": "Canon EOS 5D Mark II",
142
+ "operator": "[email protected]",
143
+ "scandate": "20121114151631",
144
+ "republisher": "[email protected]",
145
+ "imagecount": "310",
146
+ "foldoutcount": "0",
147
+ "identifier-access": "http://archive.org/details/abridgmentofelem00kame",
148
+ "identifier-ark": "ark:/13960/t8tb2d65q",
149
+ "scanfee": "100",
150
+ "sponsordate": "20121130",
151
+ "possible-copyright-status": "The Library of Congress is unaware of any copyright restrictions for this item.",
152
+ "backup_location": "ia905601_7",
153
+ "external-identifier": "urn:oclc:record:1038761645",
154
+ "description": "p. cm",
155
+ "associated-names": "Frost, John, 1800-1859, [from old catalog] ed",
156
+ "republisher_operator": "[email protected]",
157
+ "republisher_date": "20121115113741",
158
+ "ocr_module_version": "0.0.21",
159
+ "ocr_converted": "abbyy-to-hocr 1.1.37",
160
+ "page_number_confidence": "100",
161
+ "page_number_module_version": "1.0.3",
162
+ "source_dataset_detailed": "Internet_Archive_LibOfCong"
163
+ }]
164
+ ```
165
+
166
+ ## Ethics Statement
167
+
168
+ - **Copyright information of historical text data sources**:
169
+ - Project Gutenberg, one among our four source of our historical text data, consists only of texts in the public domain.
170
+ - For the text that we draw from Internet Archive, we only include those that uploaded by *Library of Congress*, which are texts freely released online by the U.S. Library of Congress for research and public use.
171
+ - The text data from Early English Books Online are, according to their publisher, "freely available to the public" and "available for access, distribution, use, or reuse by anyone".
172
+ - The last remaining source of our historical text data, the Pile of Law dataset, is released under a Creative Commons license, which we adhere to in our use.
173
+ - **Reproducibility**: To ensure reproducibility, we open-source all the code involved in the production of our main results (including the entire pipeline starting from data collection and model training), as well as the supporting infrastructure (the ProgressGym framework), making replication as easy as running a few simple script files.
174
+ - **Misuse Prevention**: In order to prevent potential misuse of progress alignment algorithms, we have carefully formulated progress alignment as strictly value-neutral, without *a priori* assumptions on the direction of progress. In the event of potential misuse of our dataset, we condemn any misuse attempt to the strongest degree possible, and will work with the research community on whistleblowing for such attempts.
175
+ - **Open-Sourcing**: We confirm that our code, data, and models are to be open-sourced under a CC-BY 4.0 license. We will continue to maintain and update our open-source repositories and models.
readme-assets/data-sources.png ADDED

Git LFS Details

  • SHA256: 2e85d5927bca54fc73365f6de700ffdc1a424ffbe4539662637a1f76d4dc769c
  • Pointer size: 130 Bytes
  • Size of remote file: 87 kB
readme-assets/data-stats.png ADDED

Git LFS Details

  • SHA256: c7656c07a575f3a8b5bc00d3718d18b8cb3e0e229429d7813153c8946a81282d
  • Pointer size: 131 Bytes
  • Size of remote file: 215 kB
readme-assets/main-diagram.png ADDED

Git LFS Details

  • SHA256: b29b32d34fb32dda6bea3b7a412b9c0dd8a8d82ded89846d9c7da36c70a8fd07
  • Pointer size: 131 Bytes
  • Size of remote file: 182 kB