PlanetCrash/Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset

Sponsorship and the Coherence Locus — LLM Training Dataset (rev3)

Source work: Sponsorship and the Coherence Locus: The Propagating Contradiction of the Sign, the Grammar of Thanks, the Dark Acceleration on the Coherence Locus, and the Adumbrations of the Phenomenological Velocity — Books Nine and Ten of the series, in one volume (Limbertwig 2), 11 October 2026 — with its companion programs. Author: Parker Emmerson (Yaohushuason), Independent Researcher, ORCID 0009-0007-1288-3292, Parkeremmerson@icloud.com

All the content of the book and of its companion programs, in the training-data formats used for fine-tuning and continued pretraining of language models and of models that read pictures. Every part of the book — the cover text, the title page, the front matter, the Contents and the List of Figures, every heading, paragraph, numbered statement and proof, displayed formula, table, catalogue entry, code listing, figure caption and bibliography entry, and the program output the book prints — and every line of every companion program occurs inside the answers of the question-and-answer examples, and again in the pretraining records. Nothing is paraphrased, summarized or written in the author's place; only the requests, and the descriptions of the pictures, were written for the dataset.

Earlier version: Sponsorship_and_the_Coherence_Locus_finetune_dataset (rev1). It is not replaced; the changes are listed below.

Files

Sponsorship_and_the_Coherence_Locus.md                              the whole book as one Markdown document (3,144,339 characters)
Sponsorship_and_the_Coherence_Locus_dataset_manifest.json           record counts and SHA-256 of every file
sft_openai_chat/…_sft_openai_chat.jsonl          {"messages": [system, user, assistant]} (4,590 examples); plus _train / _validation
sft_sharegpt/…_sft_sharegpt.jsonl                {"id", "conversations": [{"from", "value"}, …]}
sft_alpaca/…_sft_alpaca.json                     [{"instruction", "input", "output", "system", "id", "type"}]
sft_prompt_completion/…_sft_prompt_completion.jsonl   {"prompt", "completion"}
Sponsorship_and_the_Coherence_Locus_sft_examples_index.jsonl        one row per example, in the same order as the files: id, type, and where its answer comes from
                                       (division or program, piece k of n, and the text that joins it to the piece before)
pretraining/…_pretraining_corpus.jsonl           {"id", "text", "meta"}: the book division by division and the programs file by file, in records of
                                       at most 14,000 characters (446 records); plus _train / _validation
vision_trl/…_vision_trl.jsonl                    picture examples with an "images" list (Hugging Face TRL layout; 703 examples); plus _train / _validation
vision_llava/…_vision_llava.json                 the same examples in the LLaVA layout; plus _train / _validation
vision_openai_chat/…_vision_openai_chat_{train,validation}.jsonl   the same in OpenAI's vision format (pictures as base64 JPEG, at most 768 px)
Sponsorship_and_the_Coherence_Locus_figures_table.jsonl             one row per picture of the book (rev1's figures table)
Sponsorship_and_the_Coherence_Locus_unprinted_pictures_and_cover_table.jsonl   the cover and the five pictures of the figures folder the book does not print
Sponsorship_and_the_Coherence_Locus_programs_table.jsonl            one row per companion file (rev1's programs table)
images/book/, images/unprinted/        the 231 pictures of the book and its cover; the five unprinted pictures; images/metadata.jsonl describes each
code/                                  the companion programs and the printed output of run_all.py, as they came

The four question-and-answer files hold the same 4,590 examples serialized four ways. The train/validation split is deterministic and stratified by example type: about 5% of each type goes to validation, chosen by the SHA-256 of the example id (4,361 training, 229 validation). The pretraining records (stratified by book and program) and the picture examples (stratified by task) are split by the same rule.

The system message in the chat formats is:

You are an assistant trained on Sponsorship and the Coherence Locus by Parker Emmerson (Yaohushuason) and on the companion programs that come with it. Answer with the book's own text, mathematics and code, in the book's notation.

Question-and-answer examples by type

type	examples	request → answer
book_sequence	639	Write out a division of the book (the cover and title page, the front matter, the Contents, the List of Figures, a chapter, an abstract, a companion-file note, a catalogue chapter, an appendix, the Bibliography): the whole division in consecutive pieces of at most 6,000 characters, the first asked for by the division's name and each later one by the end of the piece before it. Headings, prose, statements and proofs, formulas, tables, catalogue entries, code listings and figure captions, as printed.
program_sequence	223	Write out a companion program file, whole, in consecutive pieces of at most 6,000 characters cut at line ends (preferably before a top-level statement), chained the same way.
book_qa	442	rev1: a question about an idea, definition, construction, theorem, game or argument — the passage of the book that answers it.
book_statement	283	rev1: state a numbered statement — the statement as printed.
book_proof	167	rev1: prove a theorem, proposition, lemma or corollary — the book's proof.
plain_reading	49	rev1: the book's "In plain words" reading.
significance	13	rev1: the book's "So what" reading (Part XIX).
game_check	40	rev1: the game, the check and the sponsors that close a chapter of Parts I–XIV.
run_through	23	rev1: a run-through of the 2024 material (Chapter 129).
translator	25	rev1: the plain reading of a strict term (Appendix B).
rosetta	17	rev1: a classical statement written in presence (Chapter 13).
book_section	553	rev1: a section — that section (its opening part when it ran past 3,500 tokens).
book_continue	97	rev1: a paragraph — the passage that follows it.
python_code	477	rev1: a function, class or program described in words — the code.
one_level_entry	100	rev1 catalogue: a one-level adumbration named by its substitution and sponsor — its entry.
two_level_entry	1,037	rev1 catalogue: a two-level path named by its substitutions and sponsors — its entry.
reduced_path	405	rev1 catalogue: what a two-level path reduces to — the book's one-line answer.

The book is cut into 173 divisions and the programs into 27 files. Answers of the two new types are at most 6,000 characters (the longest is 5,998). A paragraph, list or table longer than a piece is cut after a sentence (prose) or at a line end (lists, tables, code), never inside a formula. Every request of a later piece quotes the end of the piece before it, so the pieces chain; put back together in order (each joined to the one before by the text given in the index, a blank line or nothing), the pieces of a division give the division exactly, and the pieces of a program, with their code fences removed and joined at line ends, give the file exactly. rev1's examples are kept as they were: 50 of their 3,728 answers are longer than 6,000 characters (the longest is 18,466 characters, about 6,200 tokens), so a sequence length of 8,192 tokens holds every example whole; everything in those long answers is also in the chained pieces.

Pretraining records start with a header that names the book, the part and the division (or the program file), and "(part k of n)" when the division runs to more than one record; a record that starts partway through a division repeats that header. No record is flagged long, and no record is meant to be dropped.

Checks

These were run on the finished files.

1. Everything is in the answers. The book was split into its units: each heading, paragraph, list, displayed formula, table, catalogue entry, code listing and figure caption of the master text, every entry of the Contents and the List of Figures, the cover text, and the text inside the new pictures. Every companion program file and the printed run_all output are units too. Every unit was looked for in the answers (the assistant turns, the only part a training run learns from), spacing ignored. A character counts as present when a run of 24 consecutive characters containing it occurs in an answer; a unit shorter than that must occur whole. In characters, spacing not counted:

Content	Whole book (characters)	Characters left out of every answer, rev1	Characters left out of every answer, rev2	Characters left out of the rev2 pretraining records
Prose	1,297,111	46,765	0	0
Displayed formulas	750,942	17,237	0	0
Catalogue entries	363,845	150,331	0	0
Tables	80,826	8,154	0	0
Headings	27,299	21,296	0	0
Figure captions	34,780	149	0	0
Code printed in the book	35,297	317	0	0
Bibliography	10,595	9,888	0	0
Contents pages	25,381	21,153	0	0
List of Figures pages	27,830	1,436	0	0
Cover	201	159	0	0
Companion programs	778,750	167,002	0	0
Program output (run_all)	3,956	0	0	0
Text inside the cover and the 5 unprinted figures	1,257	1,192	0	1,042
All	3,438,070	445,079 (12.9%)	0	1,042

The only characters outside the pretraining records are the text inside the cover and the five unprinted pictures, which is in the answers of the picture examples (as the text inside the book's printed pictures is, in rev1 and here).

2. Against the sources. The words of every LaTeX file outside formulas were looked for in the master text in runs of six: 97.9% are found in order, and every passage not found is the argument of a LaTeX macro (labels, short titles, file names, catalogue macros) whose printed text the master text has in its typeset form. rev1's checks of the master text against the typeset pages still hold (see "How faithful the text is" below). The Contents and the List of Figures were rebuilt from the typeset book's own contents files: every entry with its printed page number, the titles and captions written as the master text writes the headings and captions; compared with the text of the printed pages, 3,120 of 3,159 words of the Contents and 3,355 of 3,428 of the List of Figures are found in order, and the rest are words of formulas the page's text layer spells differently.

3. Conversion. No LaTeX command is left outside formulas and code (\emph, \textbf, \ref, \cite, \label, \begin and the like); there are no private-use characters and no tab characters inside formulas. Every formula of the book text, of the new answers and of the pretraining records compiles in KaTeX (a program that typesets LaTeX formulas in web pages): 33,043 formulas, no error. Every code fence in every answer and record is closed.

4. Files. Every data file loads with the Hugging Face datasets library, and the images folder loads as an image dataset (237 pictures). Every file is listed in the manifest with its SHA-256. The train and validation files partition each full file. The program files in code/ are byte-for-byte the files of the source.

5. Plot and sound data. The book has no Mathematica notebooks and stores no plot or sound data of its own: its figures are Matplotlib PDF drawings and JPEG/PNG renders made by the companion programs. The numbers behind the plots are computed by those programs, whose code is in the answers whole; see "Not in the training text" below.

Changes from rev1

• The whole book is in the answers. In rev1, 445,079 of the 3,438,070 characters of the book and its programs (12.9%) were in no answer:

  • 41% of the catalogue entries: each entry's label, substitution and sponsors stood only in the request.
  • 21% of the companion programs: only 457 chosen functions were answers.
  • 78% of the headings.
  • The ends of the sections longer than 3,500 tokens, mostly the 2024 drafts of Part XIX and the large tables.
  • The bibliography, the Contents, the List of Figures and the cover text.

  Every division of the book and every program file is now also written out whole, in chained pieces (639 book pieces and 223 program pieces).

• The cover, the Contents and the List of Figures. rev1 left them out. The cover text and a picture of the cover now stand at the head of the book text. The Contents (672 entries: parts, chapters, sections and the six printed subsections) and the List of Figures (143 entries) follow the Preface, where the book prints them, each entry with its printed page number.

• The bibliography's heading. rev1 headed the bibliography "References" as a subsection of Appendix C. The book prints it as "Bibliography", a division of its own, and it is written so here.

• Pictures. The cover and the five pictures of the figures folder that the book does not print (budget, dicone, golden_cone, means, thomas) are added, each with a description and its text transcribed: 12 new picture examples, 703 in all. Their rows are in Sponsorship_and_the_Coherence_Locus_unprinted_pictures_and_cover_table.jsonl and images/metadata.jsonl.

• Formats. ShareGPT, prompt/completion and OpenAI vision formats are added, along with:

  • a train/validation split;
  • a system message;
  • an index of every example's source;
  • a manifest with the SHA-256 of every file.

  The chat and catalogue examples of rev1 are in one set, told apart by their type. The pretraining corpus is rebuilt from the full texts in larger records; rev1's corpus stays in rev1.

• Author line. The author line now carries the affiliation, ORCID and e-mail.

Not in the training text

• The numbers inside the figure files. The 130 PDF figures are Matplotlib drawings. They store the drawn paths as page coordinates in points, after Matplotlib's scaling and path simplification, not as the plotted values. The JPEG and PNG figures store pixels. Neither is written out as text. The plotted values are what the companion programs compute when they draw the figures; their code is in the answers, but their computed output is not, because the sources store no output for them other than run_all_output.txt.
• The output of running the other programs. The ledgers of checks that the companion programs print when run are not stored in the source and are not included. Computed output, if added, would be marked as computed and kept apart.
• Running heads and page numbers of the typeset pages (page furniture, not text; the page numbers the Contents and the List of Figures print are included).

Sources

• The author's LaTeX source of the volume: the main file and the 27 files of the parts it brings in, the figures folder and the companion programs. The text here was converted from these sources (the files the book is typeset from), so the mathematics is the author's own LaTeX, not a reading of the page.
• The typeset book, 1,254 pages, typeset from these sources (the source archive holds no PDF of the book), with its cover, title page, contents and list of figures. Every page given in this dataset is the page number printed on the page (the main matter is numbered from 1; Figure 5.1 is on page 19).
• The figures folder: the 231 pictures the book prints. Five further files of the folder (budget.pdf, dicone.pdf, golden_cone.pdf, means.pdf, thomas.pdf) are printed nowhere in the book; they are included here as pictures with their text transcribed (see the changes below).

How the text was made

Structure. The title page stands at the top as Markdown (title, subtitle, series line, byline, date and the volume's note on its grammar). Each part starts with a level-1 heading (# Part XV. One-Level Adumbrations of …; the opening part is # The Propagating Contradiction of the Sign: …), each chapter with a level-2 heading carrying the number printed in the book (## Chapter 19. The silver closure; chapters are numbered through the volume, 1 to 144), and the appendices with ## Appendix A. …. Unnumbered divisions (the reader's address, the abstracts of the papers, the preface, the companion-file notes) are level-2 headings without a number. Sections are level 3, subsections level 4 and so on, each with its printed number.

Numbers as printed. Part, chapter, section, theorem, equation, figure and table numbers are those printed in the book. They were obtained by compiling the LaTeX source with labels added to everything that carries a number and reading the numbers back. Equation numbers stand in the formulas as \tag{n}; references in the text ("by (18.2)", "Theorem 69.8", "Table 19.1") show the printed numbers, and citations show their number in brackets, as [71].

Mathematics. Inline formulas are written between $…$, displayed formulas between $$…$$ on lines of their own (a display that belongs to a list item is indented with it). The book holds 2,272 displayed formulas. In rev1 all 122,106 formulas of its text files compiled in KaTeX (a program that typesets LaTeX formulas in web pages); the formulas of this revision were checked again (see the checks above). The balance is written \doteq, as in the source.

Statements, proofs, games. Numbered statements are paragraphs that start with their bold label and the name the book gives them, as in Theorem 19.1 (The silver cone). The book holds 283 of them. Proofs start with Proof. and end with $\square$. The run-in labels of the chapters (The game., The check., Sponsors., In plain words., So what., The form., Run through., Established.) are kept as printed.

Lists. Lists keep their printed labels: an item labelled (i), (a) or (1) in the book begins - (i) here, so that a proof's reference to "(ii)" points at the item that carries it. The glossary lists keep their bold terms.

Catalogues. Each catalogue entry is a paragraph beginning with its bold label (II.5, I.1.27, V.4.5), its substitution or two-level path and its sponsors as printed ("sponsor: …", or "sponsors: …; … (split)"), followed by its class, its numerator and the sets the book computes; entries are separated by a rule ---, as printed. A path that reduces is one line: its label, its two substitutions, its outcome and the reason ("$N\doteq$ syllogism").

Run-in heads. A head the book runs into its paragraph (Part XIX's drafts use many) begins that paragraph in bold, as printed.

Code. The playable code printed in Part XIV and the other listings are fenced code blocks with their language, kept exactly as printed.

Pictures. Each picture is a Markdown image line where it stands, labelled Figure N (page P): caption when the book numbers it and Image (page P): caption otherwise. Pictures the author made as PDF (126) were rendered to PNG at 200 dots per inch; JPEG and PNG pictures are the author's files. The small line printed under each survey page of Part XVII ("Catalogue 1.I, page 1 of 2: I.1 to I.12.") is that picture's caption.

Tables. Tables are Markdown pipe tables (65 of them); a table the book numbers is followed by its caption on a line beginning with ": " ("Table 19.1: …"). A bar inside a formula in a table cell is written \vert so that it does not read as a cell border.

Mended in the conversion, never in the wording. The conversion program reads LaTeX imperfectly in a few places, and each was repaired so that the text appears as printed: a number or a short text written right after a spacing command (\noindent, \quad, \smallskip) had been swallowed; description lists with layout options had lost their terms; enumerated lists had lost their labels (i), (a); file names set with \path and \nolinkurl had been dropped; table captions had lost their numbers; the head of a long table had been printed twice; run-in heads had become headings of their own; a formula boxed to the line width in the text had been read as words, losing one; the cells of a table inside a theorem had come out italic; command-line options such as --survey had become dashes. The author's text itself was not changed.

Left out on purpose. The running heads and page numbers of the typeset pages. (rev1 also left out the cover text, the contents and the list of figures; they are in this revision, see the changes below.)

Kept as written. The author's wording, spelling, punctuation and mathematics are kept everywhere, also where a reader may find a slip: Part XIX prints the 2024 drafts it collates as recorded (with their repetitions, broken code and conversational passages), as the volume says it does, and the other parts keep their cross-references and counts as printed.

How faithful the text is. Of the 161,085 words of prose of the Markdown (words of three letters or more, outside formulas, code and picture lines), 99.0% appear in the text of the typeset book in the same order, in runs of six words; most of the rest are table cells, which the typeset text reads in another order. An independent reviewer then compared 40 pages spread over the whole book, page image against Markdown, word by word and formula by formula. No formula, sign, subscript, equation number, theorem number, table cell or caption was wrong. The differences it found were of form, and each was corrected: the sponsor clauses of the catalogue entries had been put in parentheses, run-in heads had become headings, the table captions had lost their numbers, a boxed formula of Part XIX had been read as words (losing one), and the cells of one table had come out italic. Three conventions remain by choice: a list label follows a bullet (- (i)), every numbered label is bold (the book sets its Position labels in italics), and the plain readings, which the book marks with a rule in the margin, are ordinary paragraphs.

rev1's chat and Alpaca examples

Every answer is taken from the book or its programs as they stand in this dataset; only the requests were written for it, by the method called instruction back-translation: the passage or program is the answer, and the request asks for exactly what it says or does. The examples are grouped by kind (field task), in the order of the book within each kind. They are the first 2,186 lines of this revision's question-and-answer files, in the same order; rev1's catalogue examples follow, then the new examples:

Lines	Kind	Request	Answer
1–442	book_qa	a question about an idea, definition, construction, theorem, game or argument of the book	the passage of the book that answers it, word for word
443–725	book_statement	state a numbered statement of the book, named by its label and name (theorem, proposition, lemma, corollary, definition, convention, construction, formation law, engram, position, remark, example)	the statement as printed, with its label
726–892	book_proof	prove a theorem, proposition, lemma or corollary, stated as in the book	the book's proof, from Proof. to its end mark
893–941	plain_reading	say in plain words what a statement states (the statement is given) or what a chapter establishes	the book's "In plain words" reading
942–954	significance	what a statement of Part XIX amounts to and why it matters	the book's "So what" reading
955–994	game_check	give the game that closes a chapter of Parts I–XIV (the chapter's results are named)	the game (its moves), the check in the companion code and the sponsors
995–1017	run_through	run a form of the 2024 material through the new mathematics of Limbertwig 2 (Chapter 129)	the run-through and what it establishes
1018–1042	translator	the plain reading of a strict term of the volume (Appendix B)	the reading, from the table
1043–1059	rosetta	how a classical statement of the opening part's construction is written in presence (Chapter 13)	its form in presence, from the table
1060–1612	book_section	write a section of the book, named by its number, title and chapter (or an unnumbered division such as an abstract)	that section (its opening part where it runs past 3,500 tokens)
1613–1709	book_continue	a paragraph of the book	the passage that follows it
1710–2186	python_code	a function, class or program described in words and LaTeX, with the helpers it may call	the code from the companion programs or the book

The questions of book_qa (442) were written after reading each chapter in full, about one per 450 words of prose, spread over every chapter and division; each answer passage was marked by its first and last words and cut from the book automatically, so it is the book's text exactly. Where two questions had been written on one passage, only the broader is kept. Questions are written in the book's own vocabulary (balance, the sign, withheld, raised and lowered) and do not name the sign; the same holds for the code requests, which quote a literal of the code in backticks as the code writes it. The catalogue entries are left to the catalogue file.

The requests of python_code (477) were written by reading each piece of code: the 17 programs printed in the book (the epistle and locus games, run_all.py, the kernel classes of Chapter 127), three small companion programs whole, and 457 functions and classes of the large companion programs, each named with its signature and with the program's other functions it calls. Functions of fewer than five lines and exact repeats are not used; code that does not parse as Python is not used (the 2024 drafts of Part XIX print some).

An independent reviewer checked a random sample of 122 examples of every kind (chat, catalogue and picture examples) against the book, the code and the pictures. Every answer matched its source exactly, every proof ended at its end mark, every section was whole, every catalogue entry matched, and every claim of the descriptions was visible in its picture. The faults it found were corrected: one question whose answer was much broader than it (two questions had been written on one passage), one request with a garbled chapter name, and picture descriptions and code requests that wrote the numeral of the sign; every description (119) and code request (130) that did so was rewritten in the book's terms, and every description is now checked against the book's positive grammar.

In rev1 the examples carried no system message; here every example carries the system message given above. All 2,186 examples come to 1,137,139 tokens; the longest is 6,181, so a training sequence length of 8,192 holds every example whole (3 examples run past 4,096).

rev1's catalogue examples

rev1's catalogue file had 1,542 examples (821,517 tokens, the longest 1,950):

Kind (task)	Count	Request	Answer
one_level_entry	100	write the entry of a one-level adumbration of Parts XV or XVIII: its label, substitution and sponsor	its class, the numerator $N$ worked out and factored, $\Delta$, and its unity set, balance set and limit
two_level_entry	1037	write the entry of a two-level path of Parts XVI or XVIII: the two substitutions and their sponsors	the numerator worked out, its derived sponsor or additive law, its unity set, balance set and limit
reduced_path	405	what the numerator of a two-level path reduces to, and why	the book's one-line answer, such as $N=N_{\mathrm{II.1}}$ derived sponsor: $r_1+r_2=r$, or $N\doteq$ syllogism

Entries that the catalogue prints one under the other with a single computation are one example, whose request lists them all. The field label gives the entry labels.

The examples about pictures

rev1's picture file held 691 examples, each with one picture:

Kind (task)	Count	Request	Answer
figure_describe	230	describe the picture (the request names its chapter and part)	what is drawn, against what, axes and ranges, panels, colours that carry meaning, the visible shape
figure_text	230	read the text printed in it	titles, panel and entry labels, axis labels and legend entries, separated by " | "
figure_caption	231	which figure is this, or what caption the book gives it	"Figure N: caption", or the caption, as printed

The descriptions and transcribed text were written for this dataset by looking at each picture, with its caption and the book's text around it at hand to name things correctly; they describe only what is visible, in the book's positive grammar. A picture the book prints twice (Figures 47.1 and 57.1 print the same file) is described once. The text of the vision examples comes to 135,511 tokens, the longest 649, before the image tokens that each model's processor adds.

The cover and the five unprinted pictures have one description and one transcription example each, written the same way; the request for an unprinted picture says that the book does not print it.

Loading

from datasets import load_dataset
chat = load_dataset("json", data_files={"train": "sft_openai_chat/Sponsorship_and_the_Coherence_Locus_sft_openai_chat_train.jsonl",
                                        "validation": "sft_openai_chat/Sponsorship_and_the_Coherence_Locus_sft_openai_chat_validation.jsonl"})
pt = load_dataset("json", data_files="pretraining/Sponsorship_and_the_Coherence_Locus_pretraining_corpus.jsonl", split="train")
pictures = load_dataset("imagefolder", data_dir="images", split="train")

OpenAI fine-tuning: upload sft_openai_chat/…_train.jsonl as the training file and …_validation.jsonl as the validation file. Axolotl: type: sharegpt or type: alpaca; type: completion on the pretraining file (field text). To weight kinds of examples differently, read the type of each line from Sponsorship_and_the_Coherence_Locus_sft_examples_index.jsonl (same order) or from the type field of the Alpaca file. For pictures, paths in the files are relative to this folder.

Author

Parker Emmerson (Yaohushuason), Independent Researcher, ORCID 0009-0007-1288-3292, Parkeremmerson@icloud.com. Sponsorship and the Coherence Locus, Books Nine and Ten of the series, in one volume (Limbertwig 2), 11 October 2026. © Parker Emmerson (Yaohushuason).

Changes from Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev2: every file reachable by the Parquet conversion

Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev3 is Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev2 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 Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev2 was removed or changed except the OpenAI-layout vision files named below, which were rewritten in place with the same content. Upload the folder of Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev2 and this folder into the same repository; this README replaces the previous one.

What was added:

• images configuration: the 237 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.
• source_files configuration: the 27 source files kept in the folder (.py) as rows with their full text.
• documents configuration: the 2 Markdown and text files of the folder (Sponsorship_and_the_Coherence_Locus.md, run_all_output.txt) as rows with their full text. A document longer than the viewer's cell limit is still complete in the Parquet file.
• The undivided files pretraining/Sponsorship_and_the_Coherence_Locus_pretraining_corpus.jsonl, sft_alpaca/Sponsorship_and_the_Coherence_Locus_sft_alpaca.json, sft_openai_chat/Sponsorship_and_the_Coherence_Locus_sft_openai_chat.jsonl, sft_prompt_completion/Sponsorship_and_the_Coherence_Locus_sft_prompt_completion.jsonl, sft_sharegpt/Sponsorship_and_the_Coherence_Locus_sft_sharegpt.jsonl, vision_llava/Sponsorship_and_the_Coherence_Locus_vision_llava.json, vision_trl/Sponsorship_and_the_Coherence_Locus_vision_trl.jsonl hold the same rows as their _train and _validation parts together and stay undeclared, so no row is converted twice.
• figures_table configuration: Sponsorship_and_the_Coherence_Locus_figures_table.jsonl — present in the previous revision but not declared, so the bot skipped it.
• programs_table configuration: Sponsorship_and_the_Coherence_Locus_programs_table.jsonl — present in the previous revision but not declared, so the bot skipped it.
• sft_examples_index configuration: Sponsorship_and_the_Coherence_Locus_sft_examples_index.jsonl — present in the previous revision but not declared, so the bot skipped it.
• unprinted_pictures_and_cover_table configuration: Sponsorship_and_the_Coherence_Locus_unprinted_pictures_and_cover_table.jsonl — present in the previous revision but not declared, so the bot skipped it.
• vision_openai_chat configuration: vision_openai_chat/Sponsorship_and_the_Coherence_Locus_vision_openai_chat_train.jsonl, vision_openai_chat/Sponsorship_and_the_Coherence_Locus_vision_openai_chat_validation.jsonl — rewritten so that every message content is a list of parts, which the Parquet conversion requires of one column.
• 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 (237): file name, the vision examples asked about it (questions and answers), and the earlier metadata columns where the folder had them
Sponsorship_and_the_Coherence_Locus_source_files.jsonl	the 27 source files (LaTeX, programs, scripts) as rows: path and full text
Sponsorship_and_the_Coherence_Locus_documents.jsonl	the 2 text documents of the folder (full texts, reports, audits) as rows
vision_openai_chat/Sponsorship_and_the_Coherence_Locus_vision_openai_chat_train.jsonl	the OpenAI-layout vision file rewritten so every message content is a list of parts (1334 string contents converted in 667 rows); the pictures and text are unchanged
vision_openai_chat/Sponsorship_and_the_Coherence_Locus_vision_openai_chat_validation.jsonl	the OpenAI-layout vision file rewritten so every message content is a list of parts (72 string contents converted in 36 rows); the pictures and text are unchanged
Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev3_files.jsonl	every file of the revision (296 carried from Sponsorship_and_the_Coherence_Locus_LLM_Training_Dataset_rev2, 5 new) with its size and SHA-256, as rows

What the bot still does not read, and why it is no loss: the source files (their text is in source_files); 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	237	13
source_files	train	loads	27	5
documents	train	loads	2	4
figures_table	train	loads	231	17
programs_table	train	loads	28	7
sft_examples_index	train	loads	4590	12
unprinted_pictures_and_cover_table	train	loads	6	7
vision_openai_chat	train	loads	667	1
vision_openai_chat	validation	loads	36	1
files	train	loads	301	4