PlanetCrash's picture
Upload folder using huggingface_hub
6258dfc verified
|
Raw History Blame Contribute Delete
14.2 kB
metadata
pretty_name: The Mirror Calculus — LLM Training Dataset (revision 2)
language:
  - en
license: other
license_name: all-rights-reserved
tags:
  - mathematics
  - logic
  - formal-grammar
  - riemann-hypothesis
  - latex
  - tikz
task_categories:
  - text-generation
  - question-answering
  - image-to-text
size_categories:
  - 1K<n<10K
configs:
  - config_name: openai_chat
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_openai_chat_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_openai_chat_validation.jsonl
  - config_name: sharegpt
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_sharegpt_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_sharegpt_validation.jsonl
  - config_name: alpaca
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_alpaca_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_alpaca_validation.jsonl
  - config_name: prompt_completion
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_prompt_completion_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_prompt_completion_validation.jsonl
  - config_name: pretraining
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_pretraining_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_pretraining_validation.jsonl
  - config_name: vision_llava
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_vision_llava_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_vision_llava_validation.jsonl
  - config_name: vision_openai
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_vision_openai_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_vision_openai_validation.jsonl
  - config_name: plate_numbers_chat
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_plate_numbers_chat_train.jsonl.gz
      - split: validation
        path: data/The_Mirror_Calculus_plate_numbers_chat_validation.jsonl.gz
  - config_name: plate_numbers_pretraining
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_plate_numbers_pretraining_train.jsonl.gz
      - split: validation
        path: data/The_Mirror_Calculus_plate_numbers_pretraining_validation.jsonl.gz
  - config_name: plate_coordinates
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_plate_coordinates_train.jsonl.gz
      - split: validation
        path: data/The_Mirror_Calculus_plate_coordinates_validation.jsonl.gz
  - config_name: source_units
    data_files:
      - split: train
        path: data/The_Mirror_Calculus_source_units_train.jsonl
      - split: validation
        path: data/The_Mirror_Calculus_source_units_validation.jsonl
  - config_name: images
    data_files:
      - split: train
        path: images/**
  - config_name: files
    data_files:
      - split: train
        path: The_Mirror_Calculus_LLM_Training_Dataset_rev2_files.jsonl

The Mirror Calculus — LLM Training Dataset, revision 2

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

This dataset carries the whole of The Mirror Calculus: A Presence-Only Mathematical Language — the edition of record deposited at doi:10.5281/zenodo.22002856 — in the forms a training run learns from. The edition of record contains, in order: the Summary of the volume (V17), the volume itself, the companion paper Deprogramming Zero, and the Rosetta Stone of Zero-Surrogate Errata. The two verification notebooks of the deposit and the text of the printed edition of 18 August 2026 (its PDF and JATS XML) are included as well.

Every answer is the book's own text. Nothing is summarized, paraphrased or written in the author's place; the questions (the user turns) only say which part of the book is wanted and where it sits.

Sources

Source file What it holds How it enters the dataset
mirror-calculus.tex The volume (10 parts, 27 chapters, 3 appendices, 305 plate drawings in TikZ) Verbatim LaTeX, cut at blank lines into 1563 units
deprogramming-zero.tex The companion paper Verbatim LaTeX, 135 units
Summary_V17.docx The Summary (V17), with 242 equations and 2 images Converted once to Markdown with TeX math; images kept as files
rosetta.json The errata: 458 transition rows and 363 pending-ruling lines Every string verbatim, grouped as in the file
verify_independence.ipynb, verify_symmetry_nogo.ipynb Two Jupyter notebooks Every cell, with every printed output after its code
mirror-calculus.xml, mirror-calculus.pdf The printed edition of 18 August 2026 Its passages whose wording the LaTeX lacks (the wording before the 111 corrections), its keywords, front matter and figure entries, the printed table of contents and the printed statement numbers
rosetta.pdf The errata as printed Its title block, table headings and footer

Questions name the part of the book by its printed numbers (Part IV, Chapter 7, Section 5.3, Theorem 3.7) as LaTeX typesets them from the source.

Files

All data files are JSON Lines. Every file is split into train and validation; the two parts of each file have no record in common, and every record belongs to exactly one of them.

Configuration Content Train records Validation records
openai_chat Question-and-answer examples, OpenAI chat format 2064 96
sharegpt Question-and-answer examples, ShareGPT format 2064 96
alpaca Question-and-answer examples, Alpaca format 2064 96
prompt_completion Question-and-answer examples, prompt/completion format 2064 96
pretraining Pretraining corpus in reading order 87 6
vision_llava Image examples, LLaVA format 297 10
vision_openai Image examples, OpenAI vision format 297 10
plate_numbers_chat Plate drawing numbers, chat format 295 10
plate_numbers_pretraining Plate drawing numbers, pretraining format 295 10
plate_coordinates Plate drawing numbers, structured 295 10
source_units Every source unit with its location 2037 96

Images: images/The_Mirror_Calculus_plate_001.png … images/The_Mirror_Calculus_plate_305.png (the 305 plate drawings rendered from their TikZ source at 450 dots per inch) and images/The_Mirror_Calculus_summary_image_1.png, images/The_Mirror_Calculus_summary_image_2.png (the two images of the Summary, as stored in the Word file).

The_Mirror_Calculus_manifest.json lists every file with its size in bytes and its SHA-256 checksum (the manifest itself excepted).

Record shapes

  • openai_chat: {"messages": [{"role": "user", ...}, {"role": "assistant", ...}]}
  • sharegpt: {"id", "conversations": [{"from": "human", ...}, {"from": "gpt", ...}]}
  • alpaca: {"id", "instruction", "input", "output"}
  • prompt_completion: {"prompt", "completion"}
  • pretraining, plate_numbers_pretraining: {"text", ...}
  • vision_llava: {"id", "image", "conversations"} with the image path relative to the dataset folder
  • vision_openai: chat messages with the image as a base64 data: URL
  • plate_coordinates: the stored numbers of every drawing statement, by type (line, circle, rectangle, arc)
  • source_units: every unit with its file, line range or cell, heading path and full text

The four question-and-answer files hold the same examples in the same order, so line n of each is the same example.

How long passages are handled

No passage is dropped for its length. A unit longer than 6,000 characters is cut into consecutive pieces — prose after a sentence, LaTeX code and tables after a line end, errata rows between rows. The first piece is asked for with the unit's question; each later piece is asked for with that question plus the last 200 characters of the piece before it. The pieces put back together reproduce the unit exactly (8 units are cut this way).

Pretraining records follow the edition's reading order (Summary, volume, Deprogramming Zero, errata, notebooks, printed edition) and are at most 14,000 characters. Each record opens with the heading path of its first passage; a record that begins partway through a cut unit first repeats the end of the piece before it.

Plate drawings

Each of the 305 plates is a TikZ drawing. Its numbers are written out as training text in their own files (plate_numbers_chat, plate_numbers_pretraining, gzip-compressed) so a trainer can weight them separately. Every drawing statement is one line, in the order of the source; coordinates are TeX points; every number is the shortest decimal that reads back to the same 64-bit value, in Wolfram notation (0., 19.8, -4.62). Example:

draw [line width 0.84, line cap round] line {{-4.62,6.6},{0.,19.8}}
fill circle {{-18.18,22.08},2.24}
draw [line width 0.84] arc {shift {-60.,3.96,3.96}, point {-22.14,0.99}, start angle -60., delta angle 300., radii {3.96,3.96}}

Each question names the drawing's place in the book and the command that drew it (\begin{tikzpicture}[baseline=-2pt] and its source line). No number in these files is computed by a tool; all are stored in the book.

The vision examples pair each rendered plate with its place in the book, its printed caption and its TikZ source. The plates carry no typeset letters: every mark is a drawn glyph of the Mirror codex. The two images of the Summary reproduce plates 10 and 20 of the volume and are paired with the Summary's captions.

Text left out

  • The author's location, printed in three places of the sources (the Summary's byline, the JATS XML's publisher and affiliation lines) and once in the Word file's document properties. It appears nowhere in this dataset.
  • README-zenodo.md, the deposit's instructions for updating the Zenodo record, which is not part of the book.

Bylines inside the book's text stay as the sources print them. Place names inside bibliography entries (publisher cities) are part of the citations and are kept.

Verification

Check Result
Whole book (characters, spacing ignored) 829,421
Characters left out of every answer, as written 4,001 — all in the Summary, where the Word file's equations and styling are written as Markdown and TeX
Characters left out of every answer, compared in a notation-neutral view (Markdown markers removed, TeX math turned into its Unicode symbols) 0
Share of the whole book left out 0.0000%
LaTeX units reproduced exactly, slice for slice yes
Cut units that reassemble exactly all 8
Plate numbers read back bit for bit 21,430 numbers in 305 drawings, 0 mismatches; a one-digit change is caught: yes
Second, independent decoding of the drawing numbers against the coordinate files 21,430 numbers, 0 mismatches
Longest answer piece / longest pretraining record 5,997 / 13,999 characters
Records shared between train and validation 0 in every file
Files loading with the Hugging Face datasets library all

The deposited PDFs of Deprogramming Zero and the Summary match, word for word, PDFs typeset again from their source files. The deposited volume PDF differs from the current LaTeX by the 111 corrections; its wording is carried by the printed-edition records. Read by the same text extractor, 67 short stretches of the deposited volume PDF match neither the retypeset PDF nor the answers word for word: 43 are two or three pieces of answer text joined by page layout (paragraph and cell boundaries), and the other 24 come from words broken by line-end hyphens, mathematical symbols the extractor spells out as letters, and table cells read across columns.

Changes from the previous revision

This is the first revision; there is no earlier revision of this dataset.

Citation

@misc{emmerson_mirror_calculus_dataset_rev1,
  author = {Parker Emmerson (Yaohushuason)},
  title  = {The Mirror Calculus — LLM Training Dataset, revision 1},
  year   = {2026},
  note   = {Built from the edition of record, doi:10.5281/zenodo.22002856}
}

Copyright

© 2026 Parker Emmerson (Yaohushuason). All rights reserved, as stated in the book.

Changes from The_Mirror_Calculus_LLM_Training_Dataset_rev1: every file reachable by the Parquet conversion

The_Mirror_Calculus_LLM_Training_Dataset_rev2 is The_Mirror_Calculus_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 The_Mirror_Calculus_LLM_Training_Dataset_rev1 was removed or changed. Upload the folder of The_Mirror_Calculus_LLM_Training_Dataset_rev1 and this folder into the same repository; this README replaces the previous one.

What was added:

  • images configuration: the 307 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.
  • 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 (307): file name, the vision examples asked about it (questions and answers), and the earlier metadata columns where the folder had them
The_Mirror_Calculus_LLM_Training_Dataset_rev2_files.jsonl every file of the revision (331 carried from The_Mirror_Calculus_LLM_Training_Dataset_rev1, 1 new) with its size and SHA-256, as rows

What the bot still does not read, and why it is no loss: 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 307 5
files train loads 332 4