PlanetCrash/Articutechnica_LLM_Training_Dataset

Articutechnica LLM Training Dataset (revision 2)

Author of the book: Parker Emmerson (Yaohushuason), Independent Researcher, ORCID 0009-0007-1288-3292, Parkeremmerson@icloud.com

This dataset turns the book Articutechnica: The Journey to the Pythagorean Operator (58 papers in nine parts, 1,217 printed pages) into training data for language models. Every piece of content in the book's source files, apart from the items listed under "Not in the training text", is placed in the form a training run learns from: the answers of question-and-answer records, plain-text pretraining records, vision records for images, and separate files for the numbers stored in the book's 3-D plot. The text is the author's own, copied exactly; nothing is summarized or paraphrased. Text that was not typed by the author (a transcription of the words inside a picture, or of a page of the two papers that survive only as PDF) is asked for as a transcription in its question, and its maker is named in the metadata: the field text_source of qa_records, transcription_by of vision_records, and composition of pretraining.

This revision replaces nothing: the previous revision, Articutechnica_Training_Dataset, stays as it was. The section "Changes from Articutechnica_Training_Dataset" lists every addition and fix.

What is in it

Configuration	One record is	Records (train / validation)
qa_openai	a question and its answer, OpenAI chat format (messages)	7,177 / 305
qa_sharegpt	the same, ShareGPT format (conversations with human and gpt turns)	7,177 / 305
qa_alpaca	the same, Alpaca format (instruction, input, output)	7,177 / 305
qa_prompt_completion	the same, prompt and completion	7,177 / 305
qa_records	the same question and answer with every metadata field (source file, line, chapter, piece number, provenance)	7,177 / 305
pretraining	a plain-text passage of at most about 14,000 characters, for reading straight through	423 / 17
vision_llava	an image with a question and its answer, LLaVA format (image path under images/)	180 / 3
vision_openai	the same, OpenAI vision format (image embedded as a data URL)	180 / 3
vision_records	the same with metadata (where the image appears, who transcribed its text)	180 / 3
plot_numbers_chat	a piece of the stored plot numbers, chat format	689 / 14
plot_numbers_pretraining	the stored plot numbers as plain text records	348 / 7
plot_numbers_computed	numbers the build computed from the plot (data ranges, box ratios), kept apart and marked computed	1 / 0

data/plot_numbers/Articutechnica_plot_coordinates.jsonl.gz holds the plot as decoded from its file: one line for each of the 44 traces, one for the layout and one for the config; it is for checking and is not declared as a training configuration.

The four question-and-answer formats hold the same records in the same order; use whichever format your trainer reads. images/ holds every picture, each named after the book. reports/ holds the check report, the inventory of the sources, and the audit of the previous revision. Articutechnica_LLM_Training_Dataset_rev1_manifest.json lists every file with its size and SHA-256. Articutechnica_LLM_Training_Dataset_rev1_build.ipynb is the executed notebook that built every file and ran every check.

The question-and-answer records by kind

Kind (subset field)	What the answers contain	Records (train / validation)	Answer characters
recitation	the LaTeX of the volume, passage by passage	2,635 / 128	3,103,383
statements	statements, definitions, proofs and abstracts	1,832 / 71	789,562
programs	the program files	16 / 0	63,379
notebooks	the two Colab notebooks, code and printed output	30 / 0	95,767
pdf_only	the two papers that survive only as PDF	34 / 2	101,790
versions	passages of earlier versions that differ from the printed text	2,627 / 104	797,680
front_matter	the author's outline and the printed contents pages	3 / 0	7,653

The 1,903 statements records hold 498 proofs (506 records, because long proofs come in several pieces), 675 named definitions, 674 named results and 48 abstracts, each asked for by name or by statement.

Loading

from datasets import load_dataset

qa = load_dataset("PlanetCrash/Articutechnica_LLM_Training_Dataset_rev1", "qa_openai")
pretraining = load_dataset("PlanetCrash/Articutechnica_LLM_Training_Dataset_rev1", "pretraining")
plot_numbers = load_dataset("PlanetCrash/Articutechnica_LLM_Training_Dataset_rev1", "plot_numbers_chat")

From a local copy, give the files directly, for example load_dataset("json", data_files={"train": "data/qa/Articutechnica_qa_openai_train.jsonl", "validation": "data/qa/Articutechnica_qa_openai_validation.jsonl"}).

How the records are made

• Recitation. Each LaTeX file of the volume (the master file, the 58 chapters and the three cover-lettering files) is cut into passages at its part, chapter, section and subsection headings; the preamble, the opening (title, author, abstract) and the bibliography are passages of their own. The answer is the passage exactly as typed, with comments, macros and labels kept. Joined in order, the answers of a file reproduce the file exactly. The one thing left out is the block of shared compatibility settings that the assembler put into each chapter so the 58 papers compile together (see "Not in the training text").
• Long texts. An answer longer than 6,000 characters is cut into consecutive pieces at the cleanest place available. In LaTeX that is, in order of preference, after a paragraph or before a heading or statement, after a line that ends a sentence, or at another line end outside displayed formulas; inside a code listing or formula that runs on for more than 6,000 characters, at a line end. In program code it is a line end at the outermost bracket level, otherwise another line end. A cut never leaves fewer than 30 characters of a paragraph on either side, and every cut falls at a line end, so some pieces of long listings end mid-statement; joined in order, the pieces reproduce the text exactly. The first piece is asked for with the question; each later piece is asked for with the same question plus the end of the piece before it, and the fields piece and pieces give its place.
• Notebooks. The two Colab notebooks exist only as PDF prints. The print keeps every code line in full, also where a line runs past the right edge of the page, so the build reads every glyph the PDF stores and places it in its column of the monospaced grid; indentation and blank lines are restored from the glyph positions, and the scattered glyphs that a page break leaves of a line are merged into that line. A scrolled output area is printed in full under the following cells, so each piece of output is given to the cell that drew it, by drawing order. Every code cell then parses as Python, and a second, separate reader of the PDFs finds the same cells, the same images and the same characters. For each cell there is a record that asks for its code and one that gives the code and asks for its printed output, in the order printed; output images are marked [output image k: file] and appear in the vision records. The prompt/completion format ends each prompt with a blank line.
• PDF-only papers. pv_bell_3.pdf (12 pages) and infinity_choiciognomy.pdf (6 pages) have no LaTeX source. Each page has a record with the PDF's own text layer (extracted with PyMuPDF; formulas lose their layout there) and a record with a LaTeX transcription of the page made by Claude (Anthropic) from the page image [1], labelled in the field text_source. The page images are vision records.
• Other versions. The uploaded originals of the chapters (before the LaTeX repairs made for printing) and four replaced uploads (the original Pipelines chapter, The Book of Loops 1.1, and two earlier uploads of The Final Pipeline) were checked unit by unit against all other answers. Every unit with any text not already present became a record of its own, so earlier wordings are kept without repeating what the printed text already has.
• Images. Every image file of the book is in a vision record: chapter figures (with their captions as written in the LaTeX), cover art, the uploaded picture final.jpg, the page preview stored in Articutechnica.pages, and the {n_nb_img} visible output images of the notebook prints (two of them are small widget graphics: a slider track and a placeholder icon). The answer gives the caption, then the text printed inside the image, transcribed by Claude (Anthropic) from the image [1]. The 3-D figure of infinity_matrix_mechanical_geometry.html is rendered in a browser and the pages of the PDF-only papers are rendered at 110 dpi; these 19 rendered images are marked computed_render.
• Plot numbers. The figure in infinity_matrix_mechanical_geometry.html (44 traces: 24 surfaces, 14 curves, 3 arrow fields and 3 sets of text labels) is written out trace by trace, every field and every number, in Wolfram notation: each number is the shortest decimal that reads back to the same 64-bit value (for example 2., 1.5*^-7, -0.), integers stay integers and empty entries are Null. Rows of a surface are one line each; long lists are cut into pieces at line ends. The first piece of a trace is asked for with the command that draws the figure, each later piece with the last line of the piece before it.
• Pretraining. The same sources as plain text records of at most about 14,000 characters, each opening with one comment line that names the book, the source and the record number. A record that starts partway through a notebook cell's output first repeats the cell's code.
• Splits. Validation holds chapters 28, 29 and 42 (the validation chapters of the previous revision) in every file, plus about 3% of the records that belong to no chapter (the master file, the notebooks, the PDF-only papers, the replaced uploads, the plot numbers), chosen by a hash of their group so the choice is the same on every build; a chained answer never straddles the two splits. Of the plot numbers, one trace (trace 38) falls in validation this way. Validation is held out from training. To train on every character of the book, train on both splits.

Changes from Articutechnica_Training_Dataset

The previous revision was audited against this revision's standard before anything was built. Under the standard's rule (a character counts as present when a run of 24 consecutive characters containing it appears in one answer), its answers left out 859,013 of the book's 2,850,072 characters (30.1%). Under the same rule this revision leaves out 0 of 2,850,072 (0.00%). The audit is in reports/Articutechnica_rev0_audit.json, with every miss listed in reports/Articutechnica_rev0_audit_misses.jsonl.gz.

Additions

1. Recitation records for every LaTeX file of the volume: 2,763 answers covering all of the text, including preambles, comments, titles and bylines, headings, statement labels, captions, tables, code listings and bibliographies, which were missing or only in questions before.
2. The 9 program files kept with the chapters and the cover (16 answers).
3. The two Colab notebook prints: 11 code cells and their printed output, including printed messages, error tracebacks and the slider and button labels (30 answers).
4. The two PDF-only papers, page by page, as text layer and as LaTeX transcription, with page images (36 answers).
5. Earlier versions: passages of the uploaded originals and of four replaced uploads that differ from the printed text (2,731 answers).
6. The author's outline in Articutechnica.pages and the contents pages of the printed volume.
7. Vision records for 183 images in LLaVA and OpenAI vision formats, with the text inside every image transcribed.
8. The plot numbers of infinity_matrix_mechanical_geometry.html in their own gzip files (chat and pretraining formats), a coordinate file, and the computed data ranges kept apart.
9. The ShareGPT, Alpaca and prompt/completion formats, beside OpenAI chat and the records file with metadata.
10. The manifest with SHA-256 checksums, the check report, the source inventory and the audit of the previous revision.

Fixes

1. No length cutoffs. Sections of all lengths are included (before: only sections of 400 to 12,000 characters), passages under 60 characters are kept, and long answers are cut into chained pieces of at most 6,000 characters instead of being skipped.
2. Pretraining records are at most about 14,000 characters (before: whole chapters, up to 341,760 characters).
3. Answers keep the source exactly: the author's macros are no longer replaced by their expansions, and \label markers are no longer removed.
4. Three silent edits of the previous revision are undone, so the answers match the printed book: the template table "Example Table" in The Book of Loops, the three headings "Figure placeholder." in Visualizing Discrete and Continuous Structures, and the coordinate −0.04 in Fractal Logical Mitosis with Expansions (see "Erratum").
5. Chapter titles, bylines and part titles are now inside answers, not only inside questions.
6. The task "write the section" is replaced by the recitation records, which cover every section; the abstract requests now ask for the abstract as written, not for a new abstract or a summary.
7. Two splits (train and validation) instead of three; the previous test chapters (17, 21 and 45) are now in train.
8. Every file is named after the book, and the card carries the author line.

Checks

All checks were run on the files as written; the full report is reports/Articutechnica_rev1_check_report.md.

Check	Result
1. Coverage of the answers (every unit of every source)	0 of 2,850,072 characters of the book left out; 0 left out across all sources including earlier versions
1. Images	183 images, each with LLaVA and OpenAI vision records; problems: 0
2. Plot numbers read back bit for bit	204,767 decimals, 196 integers, 54 Nulls; differences: 0; deliberate one-digit change caught: yes
3. Independent second decoding against the coordinate file	differences: 0
4. Files open; data files load with datasets	207 data files and pictures opened, 0 failed; train and validation disjoint in every file; every file in the manifest with its SHA-256
Notebook prints read twice by separate readers	cells 11 and 11; cells that come up short: 0; code cells that parse as Python: 11 of 11
Length rules	longest answer 6,000; longest pretraining record 13,896; none flagged "long"

Not in the training text, and why

• The 11 files in papers/Fractal_Logical_Mitosis/notebook/ (nb2_A_compass_field.py to nb3_K_hex_alpha.py). They are transcriptions of the notebook cells made from the prints while the volume was assembled, and they lost the ends of long lines and some indentation. The notebook records of this revision read the prints themselves completely (every stored glyph, indentation and blank lines restored), so these copies are left out.
• One image object in the print of "Fractal logical mitosis 2" (42 × 11,031 pixels, drawn beside the code of cell 1). Its transparency mask hides every pixel, so it shows nothing in the print. The other 77 output images are included; two of them, a slider track and a placeholder icon, are partly transparent and keep their transparency.
• The MP4 video that notebook "Fractal logical mitosis 2" saves in its last cell (compass_fractal_mitosis.mp4). The video was written to the Colab session, not to any file that was uploaded; only the print of the cell and its printed messages exist.
• The compatibility block that the assembler inserted into each printed chapter (between "% === ARTICUTECHNICA COMPAT ===" and "% === END COMPAT ==="). It is a shared set of package and macro settings written to make 58 independent papers compile in one volume; it is not the author's text. The uploaded originals, which never had it, are covered in full.
• Text that LaTeX generates while typesetting: page numbers, running heads, automatic numbers such as "Theorem 3.2", each paper's own contents page, and dates printed by \today. They are produced from the source when the PDF is compiled and are not in any source file. The words of every heading and statement are in the training text; the contents pages of the volume itself are included from the printed PDF.
• Pictures that the LaTeX source draws (68 TikZ drawings and other vector diagrams). Their complete source, including every label, is in the training text. They are not also given as separate picture files in the vision records.
• The original PDF files of the papers (56 files, 1,106 pages). Their text is the typeset form of the LaTeX sources, which are in the training text; a page-by-page comparison found no prose that the sources lack, apart from the two PDF-only papers, which are included.
• Byte-identical copies: the six Connector Theory .txt files, "The Final Pipeline 3" (identical to 3.2), duplicate image files, and the second upload of the painted cover. Each content is included once.
• Build and repair tooling of the LaTeX volume (build.sh, normalize.py, fixes/*.py, compat.tex, the folder README) and the Zenodo deposit forms. They were written to compile and deposit the volume; they are not book content.
• The Plotly JavaScript library inside infinity_matrix_mechanical_geometry.html. It is the plotting program (about 4.5 MB of library code), not data of the book; every number of the figure itself is included.
• Settings stored inside Articutechnica.pages other than the outline text (styles, calendar names, printer setting, timestamps) and its two smaller preview copies. They are application data, not writing; the outline text and the full-size page preview are included.
• The seven Google Colab HTML files. They contain only the Colab sign-in page, with no notebook content.
• The cited bibliography of the chapter "Higher Categorical Normalization and Quantum Categorical Synergies". The chapter calls \bibliography{references}, but no references.bib file was ever uploaded, so there are no entries to include.

Erratum

Chapter 39, Fractal Logical Mitosis with Expansions, gives the second nucleus of the seam at x = −0.04 in four places. A review of the volume, which recomputed the nucleus at finer resolution, found it at about −0.05 (the previous revision of this dataset had changed the text to that value). This revision keeps the text as printed.

Provenance

The chapters are the papers of Parker Emmerson (Yaohushuason), written between 2023 and 2026 and assembled into the volume in 2026. Chapters 10 and 30 are co-authored with Ryan J. Buchanan. Chapters 39 (Fractal Logical Mitosis with Expansions) and 58 (The Ratio Under a State) were drafted for the volume under the author's direction, around the author's programs and results, with the assistance of a large language model [1]; their records carry this in the field composition, as do the master file (assembled for the volume, with a draft preface) and the programs written for the volume. Transcriptions of the text inside images, and of the pages of the two PDF-only papers, were made by the same model [1] and are labelled in every record that carries them.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). Use, adaptation and redistribution, including for training models, are permitted with attribution to the author.

Citation

Parker Emmerson (Yaohushuason). Articutechnica: The Journey to the Pythagorean Operator. Training dataset, revision 1, 2026.

References

[1] Anthropic, Claude (model identifier claude-opus-5-5), large language model; used August–October 2026 for the LaTeX repair and assembly of the volume and for drafting chapters 39 and 58 under the author's direction, and 26 October 2026 for building this revision, including the transcriptions of image text and of the PDF-only pages.

Contact

Parker Emmerson (Yaohushuason), Independent Researcher, ORCID 0009-0007-1288-3292, Parkeremmerson@icloud.com

Changes from Articutechnica_LLM_Training_Dataset_rev1: every file reachable by the Parquet conversion

Articutechnica_LLM_Training_Dataset_rev2 is Articutechnica_LLM_Training_Dataset_rev1 with files added and the card's configs extended, so that the Hugging Face Parquet converter (the "parquet-converter" bot, which builds the dataset viewer and the refs/convert/parquet branch from the configurations a card declares) reaches every piece of content in the repository. No file of Articutechnica_LLM_Training_Dataset_rev1 was removed or changed. Upload the folder of Articutechnica_LLM_Training_Dataset_rev1 and this folder into the same repository; this README replaces the previous one.

What was added:

• images configuration: the 183 pictures under images/ are now an image dataset (the pictures themselves become an image column in the Parquet conversion), with images/metadata.jsonl giving each picture's vision questions and answers.
• notebook_cells configuration: the 1 .ipynb files cell by cell (68 rows), with each cell's source, its printed output and its outputs as JSON (pictures inside outputs as their base64 text).
• documents configuration: the 1 Markdown and text files of the folder (Articutechnica_rev1_check_report.md) as rows with their full text. A document longer than the viewer's cell limit is still complete in the Parquet file.
• plot_coordinates configuration: data/plot_numbers/Articutechnica_plot_coordinates.jsonl.gz — present in the previous revision but not declared, so the bot skipped it.
• audit_misses configuration: reports/Articutechnica_rev0_audit_misses.jsonl.gz — present in the previous revision but not declared, so the bot skipped it.
• reports/Articutechnica_rev1_coverage_misses.json is an empty list (no rows) and stays undeclared.
• files configuration: the manifest as a table — every file of the revision with its size and SHA-256.

New files:

File	What it holds
images/metadata.jsonl	one row per picture (183): file name, the vision examples asked about it (questions and answers), and the earlier metadata columns where the folder had them
Articutechnica_notebook_cells.jsonl	every cell of the 1 Jupyter notebooks (68 cells): source, printed output, the outputs as JSON
Articutechnica_documents.jsonl	the 1 text documents of the folder (full texts, reports, audits) as rows
Articutechnica_LLM_Training_Dataset_rev2_files.jsonl	every file of the revision (216 carried from Articutechnica_LLM_Training_Dataset_rev1, 3 new) with its size and SHA-256, as rows

What the bot still does not read, and why it is no loss: the .ipynb files (their cells are in notebook_cells); the Markdown and text files (their text is in documents); the manifest (its rows are in files). The pictures are converted through the images configuration.

Built 2026-10-30; every new configuration was loaded with the Hugging Face datasets library before this card was written (see the check table).

Loading check of the new configurations

Configuration	Split	Result	Rows	Columns
images	train	loads	183	5
notebook_cells	train	loads	68	7
documents	train	loads	1	4
plot_coordinates	train	loads	46	4
audit_misses	train	loads	21509	7
files	train	loads	219	4