alea-institute/kl3m-data-fdlp

KL3M Data: Federal Depository Library Program (FDLP)

Version 1.1: re-extracted with kaos-web

Version 1.1 replaces the text of the HTML pages in this dataset with a new extraction made with kaos-web and kaos-content. kaos-web picks the main content region of each page (readability), and kaos-content converts that region to a document tree, which is serialized as plain text. Pages are re-extracted from the same crawl captures that version 1 used.

What changed

• Re-extracted formats: HTML pages (including pages saved without a file extension, and RSS/Atom feeds served as HTML). Every other format (PDF, Word, Excel, PowerPoint, plain text, XML and other files) keeps its version 1 text unchanged and is marked text_source = "v1". Those formats will be re-extracted in a later version.
• Rows: this version has 279,088 rows: the 279,093 rows of version 1, with the same identifiers in the same order, minus 5 withheld rows (see below). No rows were added. 17,732 rows (6.4%) carry re-extracted text; 261,356 rows (93.6%) keep their version 1 text.
• Tokens: tokens was re-computed for every re-extracted row with the same tokenizer as version 1, kl3m-004-128k-cased. The tokenizer applies Unicode NFKC normalization, so decoding tokens gives the NFKC form of the extracted text (for example "…" decodes as "..." and "™" as "TM"). Version 1 behaved the same way. Every re-extracted row was checked: decoding its tokens reproduces the NFKC-normalized extraction exactly.
• score: the KL3M quality score of version 1, re-computed on the new tokens with the same KL3M code. Rows that keep version 1 text keep their version 1 score.
• 5 rows were withheld from this version because they contain a personal identification number.
• Most rows in this dataset are PDF documents; they keep their version 1 text and will be re-extracted in a later version. HTML pages were re-extracted only where the page identifier matches a re-extracted capture exactly; 85,969 HTML rows without such a match keep their version 1 text.
• Right-to-left text: 20 rows whose text is mostly in a right-to-left script (Arabic, Hebrew, Persian and similar; at least 20% of letters) are listed in flags/train_exclude.parquet (file, row, identifier). These rows stay in the dataset but are excluded from ALEA model training.

Blind review. Reviewers compared the new text with the earlier KL3M text, as cleaned for training, on 100 randomly drawn .gov pages from pages that passed the earlier training filters, without knowing which text was which. Reviewers preferred the new text 68 times and the earlier text 27 times; 5 were ties. Damage per 1,000 words:

	new text	earlier KL3M text, as cleaned for training
glued words	0.14	8.6
leftover markup	3.5	11.3
meaning-changing errors	0.10	7.8

A second blind review compared the re-extracted text with the original KL3M extraction re-run on the same captures, on 100 pages that were not pre-filtered. The pages were drawn from those re-extracted because their first extraction repeated text found across many pages of the same site (60 from six sites with heavy repeated navigation, 40 from other sites), and the new side was the first of three re-extraction rounds for these pages. Reviewers preferred the new text 90 times and the original 7 times; 3 were ties.

New columns

column	type	meaning
text_source	string	"kaos-web" for re-extracted text; "v1" for text kept from version 1
source_format	string	format of the captured source document: html, pdf, office, txt or other
p_useful	float64	advisory quality signal: estimated probability that the page is useful content rather than navigation, search facets or a stub, from quality model model_v2:b7b993065b1a on cheesecloth 0.3.0 @ 3e5167ce text features. This dataset is not filtered on it. Null for v1 rows
q_decision	string	advisory label from the same model: drop below 0.1451, review up to 0.1974, keep above, empty for empty text. It is not a filter: in a fresh blind check the 0.1451 drop threshold removed useful pages, mostly statistical tables. Null for v1 rows
region_fallback	string	which extraction path produced the text, for pages re-extracted because their first extraction matched text repeated across many pages of the same site: l3, main, readability or semantic (a kaos-web content region), bcweb (the bc-web extractor) or body (the whole page body). Null for pages kept from the first extraction; null for v1 rows
extractor	string	JSON with the kaos-web and kaos-content commit SHAs used; null for v1 rows

Known limitations

• No rows were removed for quality. p_useful and q_decision are advisory signals only.
• Pages with region_fallback = "body" carry the whole page body, which is mostly navigation and other page furniture around the content.
• Some statistical tables (for example Bureau of Labor Statistics benefit tables) still print one table cell per line.

Version 1 is kept unchanged on the v1 branch:

from datasets import load_dataset
v1 = load_dataset("alea-institute/kl3m-data-fdlp", revision="v1")

KL3M Data Project

Note: This page provides general information about the KL3M Data Project. Additional details specific to this dataset will be added in future updates. For complete information, please visit the GitHub repository or refer to the KL3M Data Project paper.

Description

This dataset is part of the ALEA Institute's KL3M Data Project, which provides copyright-clean training resources for large language models.

Dataset Details

• Format: Parquet files containing document text and metadata
• License: CC BY 4.0
• Tokenizer: The tokens field uses the kl3m-004-128k-cased tokenizer, a case-sensitive 128K vocabulary tokenizer optimized for legal, financial, and enterprise documents

Abstract

Practically all large language models have been pre-trained on data that is subject to global uncertainty related to copyright infringement and breach of contract. This creates potential risk for users and developers due to this uncertain legal status. The KL3M Data Project directly confronts this critical issue by introducing the largest comprehensive training data pipeline that minimizes risks related to copyright or breach of contract.

The foundation of this project is a corpus of over 132 million documents and trillions of tokens spanning 16 different sources that have been verified to meet the strict copyright and licensing protocol detailed in the project. We are releasing the entire pipeline, including:

1. The source code to acquire and process these documents
2. The original document formats with associated provenance and metadata
3. Extracted content in a standardized format
4. Pre-tokenized representations of the documents
5. Various mid- and post-train resources such as question-answer, summarization, conversion, drafting, classification, prediction, and conversational data

All of these resources are freely available to the public on S3, Hugging Face, and GitHub under CC-BY terms. We are committed to continuing this project in furtherance of a more ethical, legal, and sustainable approach to the development and use of AI models.

Legal Basis

This dataset is fully compliant with copyright law and contractual terms. The content is included based on the following legal foundation:

• Public domain materials
• US government works
• Open access content under permissive licenses
• Content explicitly licensed for AI training

Papers

For more information about the KL3M Data Project, please refer to:

• The KL3M Data Project: Copyright-Clean Training Resources for Large Language Models
• KL3M Tokenizers: A Family of Domain-Specific and Character-Level Tokenizers for Legal, Financial, and Preprocessing Applications

Citation

If you use this dataset in your research, please cite:

@misc{bommarito2025kl3mdata,
  title={The KL3M Data Project: Copyright-Clean Training Resources for Large Language Models},
  author={Bommarito II, Michael J. and Bommarito, Jillian and Katz, Daniel Martin},
  year={2025},
  eprint={2504.07854},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

@misc{bommarito2025kl3m,
  title={KL3M Tokenizers: A Family of Domain-Specific and Character-Level Tokenizers for Legal, Financial, and Preprocessing Applications},
  author={Bommarito II, Michael J. and Katz, Daniel Martin and Bommarito, Jillian},
  year={2025},
  eprint={2503.17247},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

About ALEA

The ALEA Institute is a non-profit research organization focused on advancing AI for business, law, and governance. Learn more at https://aleainstitute.ai/.