EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample

🔬 EAI-Taxonomy STEM w/ DCLM (100B sample)

🏆 Website | 🖥️ Code | 📖 Paper

A high-quality STEM dataset curated from web data using taxonomy-based filtering, containing 100 billion tokens of science, technology, engineering, and mathematics content.

🎯 Dataset Overview

This dataset is part of the Essential-Web project, which introduces a new paradigm for dataset curation using expressive metadata and simple semantic filters. Unlike traditional STEM datasets that require complex domain-specific pipelines, our approach leverages a 12-category taxonomy to efficiently identify and extract high-quality STEM content.

🧪 EAI-Taxonomy STEM w/ DCLM (100B tokens): Documents targeting science, engineering, medical, and computer science content that exhibit reasoning, combined with the DCLM classifier to filter for instruction-dense documents.

🏆 Performance

Our taxonomy-based approach achieves superior results with significantly less curation effort:

Dataset	MMLU-STEM	Curation Complexity
DCLM-baseline	27.7%	General web filtering
FineWeb-Edu	26.7%	Educational filtering
EAI-Taxonomy STEM	29.1%	Simple semantic filter
EAI-Taxonomy STEM w/ DCLM	34.5%	+ DCLM classifier

Results show +24.5% improvement over DCLM and +29.2% improvement over FineWeb-Edu.

🔍 Key Findings

• Strong STEM Performance: Outperforms baseline and educational datasets beyond standard error
• Efficient Curation: Achieves superior results without complex domain-specific pipelines
• Broad Coverage: Encompasses science, engineering, medical, and computer science domains
• Quality Focus: Selects high-quality document types and filters for reasoning content

Dataset Schema Documentation

Overview

This dataset contains web-crawled text data with comprehensive metadata, quality signals, and taxonomic classifications. Each record represents a document extracted from web archives with detailed provenance tracking and quality assessment metrics.

Core Fields

Field	Type	Description	Path
id	Int64	Unique identifier based on document hash	id
text	String	The main textual content of the document	text

EAI Taxonomy Classification

Comprehensive hierarchical classification system with primary and secondary labels - the most important feature of this dataset. The taxonomy is designed to provide detailed subject categorization, document type identification, content quality assessment, and extraction quality indicators.

Free Decimal Correspondence (FDC)

A Dewey Decimal-inspired classification system with 3-level hierarchical labels. The FDC provides nested categories where each successive level refines its parent category. It's designed to be compatible with the Dewey Decimal System for library cataloging.

Level Structure:

• Level 1: Top-level categories (0-9) covering broad subject areas like General works, Philosophy, Religion, Social Sciences, etc.
• Level 2: Sub-divisions (00-99) that refine Level 1 categories
• Level 3: Specific categories (000-999) that further refine Level 2 categories

Component	Description	Path
Primary Code	Main classification code	eai_taxonomy.free_decimal_correspondence.primary.code
Primary Level 1	Top-level category (0=General works, 1=Philosophy, 2=Religion, 3=Social Sciences, 4=Language, 5=Science, 6=Technology, 7=Arts, 8=Literature, 9=History/Geography)	eai_taxonomy.free_decimal_correspondence.primary.labels.level_1
Primary Level 2	Mid-level category	eai_taxonomy.free_decimal_correspondence.primary.labels.level_2
Primary Level 3	Specific category	eai_taxonomy.free_decimal_correspondence.primary.labels.level_3
Secondary Code	Alternative classification code	eai_taxonomy.free_decimal_correspondence.secondary.code
Secondary Level 1	Alternative top-level category	eai_taxonomy.free_decimal_correspondence.secondary.labels.level_1
Secondary Level 2	Alternative mid-level category	eai_taxonomy.free_decimal_correspondence.secondary.labels.level_2
Secondary Level 3	Alternative specific category	eai_taxonomy.free_decimal_correspondence.secondary.labels.level_3

We recommend this viewer for easily navigating the FDC categories when curating filters: https://www.librarything.com/mds

Bloom's Taxonomy Integration

Based on Anderson and Krathwohl's 2001 revision of Bloom's Taxonomy of Educational Objectives, providing two complementary categorization dimensions for educational content analysis.

Knowledge Domain

Categorizes the type of knowledge demonstrated in the document:

Component	Description	Path
Primary Code	Main knowledge domain code	eai_taxonomy.bloom_knowledge_domain.primary.code
Primary Label	Main knowledge domain label	eai_taxonomy.bloom_knowledge_domain.primary.label
Secondary Code	Alternative knowledge domain code	eai_taxonomy.bloom_knowledge_domain.secondary.code
Secondary Label	Alternative knowledge domain label	eai_taxonomy.bloom_knowledge_domain.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
1	Factual	Basic elements to learn or solve problems
2	Conceptual	Interrelationships between basic elements within larger context
3	Procedural	Methods and techniques in the discipline
4	Metacognitive	Awareness of how learning works in relation to oneself

Cognitive Processing Level

Assesses the learning and thinking skill levels demonstrated by the document author:

Component	Description	Path
Primary Code	Main cognitive process code	eai_taxonomy.bloom_cognitive_process.primary.code
Primary Label	Main cognitive process label	eai_taxonomy.bloom_cognitive_process.primary.label
Secondary Code	Alternative cognitive process code	eai_taxonomy.bloom_cognitive_process.secondary.code
Secondary Label	Alternative cognitive process label	eai_taxonomy.bloom_cognitive_process.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
1	Remember	Retrieve relevant knowledge from memory
2	Understand	Determine meaning of instructional messages
3	Apply	Use a procedure in a given situation
4	Analyze	Break materials into components and determine relationships
5	Evaluate	Make judgments based on criteria and standards
6	Create	Create new or original work

Document Characteristics

Document Type v1

In-house classification of common web document types and formats:

Component	Description	Path
Primary Code	Main document type code	eai_taxonomy.document_type_v1.primary.code
Primary Label	Main document type label	eai_taxonomy.document_type_v1.primary.label
Secondary Code	Alternative document type code	eai_taxonomy.document_type_v1.secondary.code
Secondary Label	Alternative document type label	eai_taxonomy.document_type_v1.secondary.label

Possible Values:

Code	Label	Examples
-1	Abstain	Unable to classify
1	News/Editorial	CNN articles, opinion columns
2	Academic/Research	ArXiv papers, research articles
3	Reference/Encyclopedic/Educational	FAQs, Wikipedia entries
4	Code/Software	GitHub repos, code examples
5	Social/Forum	Conversation threads, Q&A boards
6	Promotional/Advertisement	Product pages, calls to action
7	Search/Directory/Bibliography	Link pages, search results
8	Adult/Pornographic	Adult content
9	Personal/Misc	Blogs, user profiles
10	Machine-Generated	Lorem ipsum, garbled text
11	Legal/Regulatory	Contracts, terms of service
12	Government/Political	Legislation, press releases
13	Literary/Creative	Poems, short stories
14	Reviews/Critiques	Film critiques, product reviews
15	E-Commerce/Marketplace	eBay listings, Amazon pages
16	Images/Videos/Audio	YouTube videos, Imgur pages
17	Other/Unclassified	Documents that resist classification

Document Type v2

Updated classification based on WebOrganizer taxonomy with refined categories for improved document classification accuracy:

Component	Description	Path
Primary Code	Main document type code (v2)	eai_taxonomy.document_type_v2.primary.code
Primary Label	Main document type label (v2)	eai_taxonomy.document_type_v2.primary.label
Secondary Code	Alternative document type code (v2)	eai_taxonomy.document_type_v2.secondary.code
Secondary Label	Alternative document type label (v2)	eai_taxonomy.document_type_v2.secondary.label

Complete Value Mapping:

Code	Label	Examples
-1	Abstain	Documents requiring human review
1	About (Org.)	Company about pages, mission statements
2	About (Personal)	Personal bios, LinkedIn profiles
3	Academic Writing	Research papers, abstracts, dissertations
4	Audio Transcript	Interview transcripts, court records, captions
5	Comment Section	Reddit threads, blog comments
6	Content Listing	Site maps, product catalogs, directory listings
7	Creative Writing	Song lyrics, novel excerpts, poetry
8	Documentation	API docs, README files, user manuals
9	FAQ	FAQ pages, Q&A lists
10	Knowledge Article	Wikipedia articles, Britannica entries
11	Legal Notices	Privacy policies, license agreements, terms of service
12	Listicle	Buzzfeed-style articles, "Top 10" lists
13	News (Org.)	Government blog posts, corporate announcements
14	News Article	Newspaper articles, CNN content, breaking news
15	Nonfiction Writing	Editorials, obituaries, memoirs, opinion pieces
16	Personal Blog	Personal journals, diary entries, lifestyle blogs
17	Product Page	Product descriptions, course offerings, sales pages
18	Q&A Forum	Quora posts, Stack Exchange discussions
19	Spam / Ads	SEO keyword stuffing, promotional spam
20	Structured Data	Datasheets, glossaries, JSON files, databases
21	Customer Support	Help articles, troubleshooting guides
22	Truncated	Paywalled sites, image galleries, partial content
23	Tutorial	Cooking recipes, WikiHow pages, step-by-step guides
24	User Review	Yelp reviews, TripAdvisor feedback, product reviews
25	Other/Unclassified	Miscellaneous documents not fitting other categories

Extraction Artifacts

Assessment of technical extraction quality, identifying issues from HTML-to-text conversion:

Component	Description	Path
Primary Code	Main extraction artifact code	eai_taxonomy.extraction_artifacts.primary.code
Primary Label	Main extraction artifact label	eai_taxonomy.extraction_artifacts.primary.label
Secondary Code	Alternative extraction artifact code	eai_taxonomy.extraction_artifacts.secondary.code
Secondary Label	Alternative extraction artifact label	eai_taxonomy.extraction_artifacts.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
0	No Artifacts	Clean text with no leftover HTML or irrelevant elements
1	Leftover HTML	HTML/code artifacts remaining after extraction
2	Text Extraction Errors	Broken math expressions, encoding errors, improperly parsed tables
3	Irrelevant Content	Headers, footers, nav menus extracted by mistake
4	Indeterminate	Insufficient content to judge

Missing Content

Assessment of content completeness and extraction success:

Component	Description	Path
Primary Code	Main missing content code	eai_taxonomy.missing_content.primary.code
Primary Label	Main missing content label	eai_taxonomy.missing_content.primary.label
Secondary Code	Alternative missing content code	eai_taxonomy.missing_content.secondary.code
Secondary Label	Alternative missing content label	eai_taxonomy.missing_content.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
0	No Missing Content	Complete and coherent text
1	Truncated Snippets	Obvious "...", incomplete paragraphs, cut-off text
2	Click Here References	"Download here", "Click here" without linked content
3	Incoherent Flow	Unreadable or illogical flow due to missing context
4	Missing Images or Figures	Placeholders or references to missing visual content
5	Missing Referenced Data	References to absent tables/datasets (e.g., "See Table 3")
6	Indeterminate	Insufficient content to judge

Text Structure Information

Field	Type	Description	Path
Line Start Indices	List[Int32]	Starting indices of each line	line_start_n_end_idx.line_start_idx
Line End Indices	List[Int32]	Ending indices of each line	line_start_n_end_idx.line_end_idx

Content Quality Dimensions

Quality assessment inspired by NaturalReasoning and FineWeb efforts to categorize web data by information sophistication.

Reasoning Depth

Assesses the complexity and sophistication of logical reasoning in the document:

Component	Description	Path
Primary Code	Main reasoning depth code	eai_taxonomy.reasoning_depth.primary.code
Primary Label	Main reasoning depth label	eai_taxonomy.reasoning_depth.primary.label
Secondary Code	Alternative reasoning depth code	eai_taxonomy.reasoning_depth.secondary.code
Secondary Label	Alternative reasoning depth label	eai_taxonomy.reasoning_depth.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
1	No Reasoning	Facts present but no evidence of reasoning
2	Basic Reasoning	Basic analysis with minimal explanation and summarization
3	Intermediate Reasoning	Some logical steps connecting ideas and structured thinking
4	Advanced Reasoning	Multi-step reasoning and thorough analysis with well-developed explanations
5	Exceptional Reasoning	Novel abstractions, theoretical frameworks, long chain-of-thought, original insights, or proofs
6	Indeterminate	Insufficient context to judge

Technical Correctness

Evaluates the accuracy and precision of technical information:

Component	Description	Path
Primary Code	Main technical correctness code	eai_taxonomy.technical_correctness.primary.code
Primary Label	Main technical correctness label	eai_taxonomy.technical_correctness.primary.label
Secondary Code	Alternative technical correctness code	eai_taxonomy.technical_correctness.secondary.code
Secondary Label	Alternative technical correctness label	eai_taxonomy.technical_correctness.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
1	Technically Flawed	Significant errors undermining content validity
2	Partially Correct	Some correctness but contains flaws, omissions, or errors
3	Mostly Correct	Technical correctness with minor flaws or incomplete explanations
4	Highly Correct	High technical correctness with precise definitions and clear explanations
5	Exceptionally Correct	Exceptional technical correctness with formal proofs and flawless content
6	Not Applicable/Indeterminate	No technical content or insufficient context

Education Level

Assesses the appropriate educational background required to comprehend the content:

Component	Description	Path
Primary Code	Main education level code	eai_taxonomy.education_level.primary.code
Primary Label	Main education level label	eai_taxonomy.education_level.primary.label
Secondary Code	Alternative education level code	eai_taxonomy.education_level.secondary.code
Secondary Label	Alternative education level label	eai_taxonomy.education_level.secondary.label

Possible Values:

Code	Label	Description
-1	Abstain	Unable to determine
1	General Audience	Accessible to anyone with basic literacy; simple terms
2	High School Level	Requires high school education; specialized terminology explained for non-experts
3	Undergraduate Level	Requires college education; uses specialized terminology and assumes background knowledge
4	Graduate/Expert Level	Requires graduate education or domain expertise; assumes deep background knowledge
5	Indeterminate	Insufficient content to judge educational level

Metadata

Metadata Structure

The metadata field contains a nested structure with web archive information:

Field	Type	Description	Path
URL Information
URL	String	Original URL of the document	metadata.url
Source Domain	String	Domain name of the source	metadata.source_domain
Snapshot ID	String	Identifier for the web archive snapshot	metadata.snapshot_id
WARC Metadata		WARC (Web ARChive) format metadata
Content Length	String	Size of the content	metadata.warc_metadata.Content-Length
Content Type	String	MIME type of the content	metadata.warc_metadata.Content-Type
Block Digest	String	Checksum of the WARC block	metadata.warc_metadata.WARC-Block-Digest
Concurrent To	String	Related WARC records	metadata.warc_metadata.WARC-Concurrent-To
Date	String	Timestamp of the crawl	metadata.warc_metadata.WARC-Date
IP Address	String	Source server IP address	metadata.warc_metadata.WARC-IP-Address
Payload Type	String	Identified content type	metadata.warc_metadata.WARC-Identified-Payload-Type
Payload Digest	String	Checksum of the payload	metadata.warc_metadata.WARC-Payload-Digest
Record ID	String	Unique WARC record identifier	metadata.warc_metadata.WARC-Record-ID
Target URI	String	Original target URL	metadata.warc_metadata.WARC-Target-URI
Truncated	String	Truncation status	metadata.warc_metadata.WARC-Truncated
Type	String	WARC record type	metadata.warc_metadata.WARC-Type
Warcinfo ID	String	Associated warcinfo record	metadata.warc_metadata.WARC-Warcinfo-ID
Additional Info
WARC Info	String	Additional WARC information	metadata.warc_info

Quality Signals

The dataset includes two comprehensive quality assessment frameworks:

Red Pajama v2 Quality Metrics

Text quality indicators derived from the Red Pajama v2 filtering pipeline:

Content Structure Metrics

Metric	Description	Path
Original Length	Original document length	quality_signals.red_pajama_v2.ccnet_original_length
Original Lines	Number of lines in original document	quality_signals.red_pajama_v2.ccnet_original_nlines
Sentence Count	Total sentence count	quality_signals.red_pajama_v2.rps_doc_num_sentences
Word Count	Total word count	quality_signals.red_pajama_v2.rps_doc_word_count
Mean Word Length	Average word length	quality_signals.red_pajama_v2.rps_doc_mean_word_length

Language Quality Metrics

Metric	Description	Path
Stop Word Fraction	Proportion of stop words	quality_signals.red_pajama_v2.rps_doc_stop_word_fraction
Unique Words Fraction	Fraction of unique words	quality_signals.red_pajama_v2.rps_doc_frac_unique_words
All Caps Words	Fraction of words in all capitals	quality_signals.red_pajama_v2.rps_doc_frac_all_caps_words
Non-Alphabetic Words	Fraction of non-alphabetic words	quality_signals.red_pajama_v2.rps_doc_frac_no_alph_words
Unigram Entropy	Entropy measure of word distribution	quality_signals.red_pajama_v2.rps_doc_unigram_entropy

Content Pattern Analysis

Metric	Description	Path
Curly Bracket Density	Curly bracket density (code indicator)	quality_signals.red_pajama_v2.rps_doc_curly_bracket
Symbol-to-Word Ratio	Symbol-to-word ratio	quality_signals.red_pajama_v2.rps_doc_symbol_to_word_ratio
Ellipsis Line Endings	Lines ending with ellipsis	quality_signals.red_pajama_v2.rps_doc_frac_lines_end_with_ellipsis
Lorem Ipsum Detection	Lorem ipsum text detection	quality_signals.red_pajama_v2.rps_doc_lorem_ipsum
Offensive Content	Potentially offensive content detection	quality_signals.red_pajama_v2.rps_doc_ldnoobw_words
UT1 Blacklist	UT1 blacklist filtering score	quality_signals.red_pajama_v2.rps_doc_ut1_blacklist

Duplication Detection

Metric	Description	Path
5-gram Duplication	Character-level duplication for 5-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_5grams
6-gram Duplication	Character-level duplication for 6-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_6grams
7-gram Duplication	Character-level duplication for 7-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_7grams
8-gram Duplication	Character-level duplication for 8-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_8grams
9-gram Duplication	Character-level duplication for 9-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_9grams
10-gram Duplication	Character-level duplication for 10-grams	quality_signals.red_pajama_v2.rps_doc_frac_chars_dupe_10grams
Top 2-gram Coverage	Most frequent 2-gram coverage	quality_signals.red_pajama_v2.rps_doc_frac_chars_top_2gram
Top 3-gram Coverage	Most frequent 3-gram coverage	quality_signals.red_pajama_v2.rps_doc_frac_chars_top_3gram
Top 4-gram Coverage	Most frequent 4-gram coverage	quality_signals.red_pajama_v2.rps_doc_frac_chars_top_4gram

Domain Importance Scores

Metric	Description	Path
Books Importance	Similarity to book content	quality_signals.red_pajama_v2.rps_doc_books_importance
Books Importance (Length Corrected)	Length-corrected books similarity	quality_signals.red_pajama_v2.rps_doc_books_importance_length_correction
OpenWebText Importance	Similarity to OpenWebText	quality_signals.red_pajama_v2.rps_doc_openwebtext_importance
OpenWebText Importance (Length Corrected)	Length-corrected OpenWebText similarity	quality_signals.red_pajama_v2.rps_doc_openwebtext_importance_length_correction
Wikipedia Importance	Similarity to Wikipedia	quality_signals.red_pajama_v2.rps_doc_wikipedia_importance
Wikipedia Importance (Length Corrected)	Length-corrected Wikipedia similarity	quality_signals.red_pajama_v2.rps_doc_wikipedia_importance_length_correction

FastText Classification Scores

Domain and content type classification probabilities:

Metric	Description	Path
DCLM Score	DataComp-LM classifier score	quality_signals.fasttext.dclm
English Confidence	English language confidence	quality_signals.fasttext.english
Educational Content	Educational content approximation	quality_signals.fasttext.fineweb_edu_approx
General Math	General mathematics content	quality_signals.fasttext.eai_general_math
Web Math	OWM Web-based mathematics content	quality_signals.fasttext.eai_open_web_math
Code Content	Code content detection	quality_signals.fasttext.eai_web_code

How to Load the Dataset

This section provides examples of how to load the EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample dataset using different Python libraries and frameworks.

Using Hugging Face Datasets (Standard Method)

The simplest way to load the dataset is using the Hugging Face datasets library:

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample")

# View dataset structure
print(dataset)
print(f"Number of examples: {len(dataset['train'])}")

You can also load the dataset in streaming mode to avoid downloading the entire dataset at once:

from datasets import load_dataset

# Load in streaming mode
dataset = load_dataset("EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample", streaming=True)
data_stream = dataset["train"]

# Iterate through examples
for example in data_stream.take(5):
    print(example)

Using PySpark

For large-scale distributed processing, you can load the dataset using PySpark with the pyspark_huggingface library:

# First install the required library:
# pip install pyspark_huggingface

import pyspark_huggingface
from pyspark.sql import SparkSession

# Initialize Spark session
spark = SparkSession.builder.appName("EAI-Taxonomy-STEM-w-DCLM").getOrCreate()

# Load the dataset using the "huggingface" data source
df = spark.read.format("huggingface").load("EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample")

# Basic dataset exploration
print(f"Dataset shape: {df.count()} rows, {len(df.columns)} columns")
df.show(10)
df.printSchema()

# Load only specific columns for efficiency
df_subset = (
    spark.read.format("huggingface")
    .option("columns", '["column1", "column2"]')  # Replace with actual column names
    .load("EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample")
)

# Run SQL queries on the dataset
df.createOrReplaceTempView("eai_taxonomy_stem_w_dclm_dataset")
result = spark.sql("""
    SELECT COUNT(*) as total_examples
    FROM eai_taxonomy_stem_w_dclm_dataset
""")
result.show()

Using Daft

Daft provides a modern DataFrame library optimized for machine learning workloads. You can load the dataset directly from Hugging Face:

import daft

# Load the entire dataset
df = daft.read_parquet("hf://datasets/EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample")

# Basic exploration
print("Dataset schema:")
df.schema()

print("First 5 rows:")
df.show(5)

If you need to access private datasets or use authentication:

import daft
from daft.io import IOConfig, HTTPConfig

io_config = IOConfig(http=HTTPConfig(bearer_token="your_token"))
df = daft.read_parquet("hf://datasets/EssentialAI/eai-taxonomy-stem-w-dclm-100b-sample", io_config=io_config)

Installation Requirements

Make sure you have the required libraries installed:

# For Hugging Face datasets
pip install datasets

# For PySpark with Hugging Face integration
pip install pyspark_huggingface

# For Daft
pip install daft

📜 License

Essential-Web-v1.0 contributions are made available under the ODC attribution license; however, users should also abide by the Common Crawl - Terms of Use. We do not alter the license of any of the underlying data.

📝 Citation

@misc{ai2025essentialwebv1024ttokens,
      title={Essential-Web v1.0: 24T tokens of organized web data}, 
      author={Essential AI and : and Andrew Hojel and Michael Pust and Tim Romanski and Yash Vanjani and Ritvik Kapila and Mohit Parmar and Adarsh Chaluvaraju and Alok Tripathy and Anil Thomas and Ashish Tanwer and Darsh J Shah and Ishaan Shah and Karl Stratos and Khoi Nguyen and Kurt Smith and Michael Callahan and Peter Rushton and Philip Monk and Platon Mazarakis and Saad Jamal and Saurabh Srivastava and Somanshu Singla and Ashish Vaswani},
      year={2025},
      eprint={2506.14111},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.14111}, 
}