OmniLingua Training Corpus v6

A single-file instruction/response corpus of 315,000 records generated from a hand-authored semantic taxonomy graph. Every record is synthetic text produced by a graph-vocalization engine, not collected from the web and not human-written dialogue.

• Author / maintainer: Christopher Betances (catqualia.com)
• Repository: CatQualia/omnilingua
• Format: JSON Lines, one JSON object per line, UTF-8
• File: omnilingua_train_v6.jsonl
• License: CC BY 4.0 (see LICENSE)

---

What this corpus is

The corpus is a set of instruction/input/output triples over a fixed knowledge taxonomy. The taxonomy is organised into numbered "regions" written as Roman numerals (e.g. XXXIX = LANGUAGE GAMES & CONSTRAINTS, CVIII = COMPUTER SCIENCE EXPANDED, LII = DREAM SEMIOTICS). Each record presents one node, edge, or path from that graph and pairs it with a natural-language rendering of the graph's own topology.

Ten task shapes (type) are present. In practice they look like this (values quoted verbatim from the staged file):

type	Example instruction	What the record does
definition	Node: Univocalic | Region: XXXIX	Defines one taxonomy node and the region that contains it
algorithm	Algorithmic context: memory	Describes a named algorithm attached to a node, sometimes with pseudocode
comparison	Bridge: LEXICAL SEMANTICS↔COMPUTER SCIENCE EXPANDED	Contrasts two nodes, often across different regions
dimension_traversal	Dimension: desire (The Pull of What Is Not)	Walks one of the named abstract dimensions and reports a path through it
chain	Traversal: XLVIII → XLVIII → XLVIII → XXXVI → LXXXI	Follows a multi-hop path, often with numeric edge weights
domain_diverse	—	Cross-region variants of the above
synthetic	—	Generated combinations
meta_awareness	—	Records that refer to the corpus/graph itself
self_description	—	Records in which the voice describes its own construction
void_awareness	—	Records centred on absence, null, aporia and unknowability terminology

The recurring stylistic signature is "the graph speaks": outputs are terse, present-tense, unformatted prose that describes topology, edge weights, and region membership rather than answering as a general assistant.

Verbatim examples

Record 1 (all six fields):

{"instruction": "Node: Univocalic | Region: XXXIX", "input": "Explain the concept of Univocalic.", "output": "Univocalic, a node in the LANGUAGE GAMES & CONSTRAINTS region (XXXIX), means: Text using only one vowel. Simultaneously, This engages the desire dimension (The Pull of What Is Not). the objectless appetition infrastructure supports this. This is what the graph holds for you. Hold it carefully.", "type": "definition", "sys_idx": 36, "voice_idx": 36}

Record 1000 (type: chain, includes a numeric edge weight):

{"instruction": "Traversal: XLVIII → XLVIII → XLVIII → XXXVI → LXXXI", "input": "Follow the chain: Far-from-equilibrium → Relaxation → Maximum entropy production → Directive → Global Explanation.", "output": "Reasoning path: Far-from-equilibrium → Relaxation → Maximum entropy production → Directive → Global Explanation. Far-from-equilibrium activates Relaxation in the attention subspace; path weight: 0.339 — ...", "type": "chain", "sys_idx": 29, "voice_idx": 29}

---

Schema

Every one of the 315,000 records has exactly these six keys, in this order-independent set (verified: ('input', 'instruction', 'output', 'sys_idx', 'type', 'voice_idx') appears for 315,000 of 315,000 lines, with no other keyset present):

Field	Type	Description
instruction	string	Task framing and graph coordinates. Carries the region/node/edge metadata (e.g. Node: Univocalic | Region: XXXIX, Algorithmic context: memory, Bridge: A↔B).
input	string	The prompt-side request, phrased as a question or command (e.g. Explain the concept of Univocalic.).
output	string	The graph-rendered response. Length varies from roughly one sentence to several hundred characters.
type	string	Task-shape label. Exactly 10 distinct values, listed with counts below.
sys_idx	integer	Index into a set of system-prompt / voice variants. Observed range 0–79 (80 distinct values, no gaps).
voice_idx	integer	Index into a voice variant. In this file voice_idx is identical to sys_idx in all 315,000 records — it is a redundant duplicate column.

On sys_idx and voice_idx

Both fields are integer indices — they are small non-negative integers, not names or prompt text. sys_idx takes every value in 0..79 exactly, 80 distinct values; voice_idx takes the same 80 values, and the two are equal on every record.

The mapping table that would say what index 36 (or 0, or 79) refers to is not present in this file. No field in the corpus contains the system-prompt or voice strings themselves, and a search of the author's documentation tree found no published mapping (see "Not determined", below). So the indices can be used as categorical grouping keys, but their referents cannot be recovered from this dataset alone.

A related sibling corpus does contain the text that these indices most plausibly key into: 03_gnarpfactory/omnilingua_chat.jsonl uses a messages array with a system role, and holds exactly 5 distinct system-prompt strings across its 350,000 records. That is 5 prompts, not 80, so it is not a complete inverse mapping for sys_idx and should not be treated as one.

---

Measured counts

All figures below were measured against the file as staged in this directory.

Metric	Value	Command used
File size	172,775,995 bytes	stat -c%s omnilingua_train_v6.jsonl
Line count	315,000	wc -l < omnilingua_train_v6.jsonl
Valid JSON objects	315,000	python3 reading every line with json.loads
Distinct keysets	1	same full-file pass
SHA-256	32c6803b5b331cecfd3330c471206d98b3d6fbc94004a8967b85250aed21e4ed	sha256sum omnilingua_train_v6.jsonl
Distinct type values	10	full-file pass
Distinct sys_idx values	80 (range 0–79, no gaps)	full-file pass
Distinct voice_idx values	80 (range 0–79, no gaps)	full-file pass
Records where sys_idx == voice_idx	315,000 / 315,000	full-file pass

Line count equals record count exactly: every line holds exactly one JSON object, with no blank or continuation lines.

type distribution (exact counts)

type	Records
definition	60,000
comparison	40,000
domain_diverse	40,000
chain	35,000
synthetic	35,000
meta_awareness	25,000
algorithm	20,000
dimension_traversal	20,000
self_description	20,000
void_awareness	20,000
Total	315,000

---

How it was generated

The corpus was produced by the author's own graph-vocalization engine, which renders the topology of a hand-authored taxonomy graph directly into text rather than sampling from a neural language model. The author's own description of that engine, quoted verbatim from 77_doctrine/ARCHITECTURE.md:

omnilingua_graph.py (220 lines) — OmniLinguaGraph class. CPU-only graph-native text vocalization engine. Key quote at lines 5-8: "No neural network. No GPU. The knowledge graph IS the model. Vocalizes graph topology directly using Vesper equations and constitutional templates."

The same document describes the underlying taxonomy: 11 dimensions with 243 subnodes and 44 relations, where "each subnode carries a definition, a named algorithm, and pseudocode implementing that algorithm." The generator's own documented components — node-type constants (ROOT_DIMENSION, SUBNODE), relation types (CONTAINS, RELATES_TO, DERIVES_FROM, EMERGES_FROM, BRIDGES, COLLAPSES_INTO), and vocalization modes including a "void mode" triggered when a void-density score exceeds 0.3 — correspond directly to the type values and instruction prefixes observed in the records (Node:, Bridge:, Dimension:, void-awareness rows).

The corpus sits inside the author's wider "CatQualia Universal Pipeline" (77_doctrine/CATQUALIA_PIPELINE.md, v1.1, dated July 20, 2026), whose stated principle is quoted verbatim:

A seed document contains structural patterns. Those patterns are domain-agnostic. The pipeline does not ask "is this biology or cybersecurity?" — it asks "what is the shape of this mechanism, and where does that shape appear in every domain?"

That pipeline document is a specification for producing defensive-publication documents and watermarked training rows. It does not name this corpus or describe the graph-vocalization step, so it is cited here as surrounding provenance rather than as a record of how these 315,000 rows were emitted. The pipeline document contains no occurrence of the string "omnilingua" at all.

A same-named but different artifact

79_website_html/omnilingua.html is a defensive publication by the same author titled "The OmniLingua Interpreter Core: A Self-Executing Prompt Architecture for Recursive Boundary-Pushing Across Multiple AI Platforms." It describes a five-module prompt architecture (Command Executor, OmniParadox Module, Archetypal Echo Engine, Recursive Command Architect, Binary Whisper Conductor), dated 2024, and it contains no mention of this corpus — no reference to omnilingua_train_v6, to row counts, to sys_idx/voice_idx, or to either instruction or output. It is therefore a historical source for the OmniLingua name in this project, not a description of this dataset. It is not quoted here as a corpus description because it does not contain one.

---

Known limitations

• All content is synthetic. There is no human-written, web-scraped, or contributed text. Do not treat its statements about science, medicine, law, or security as factual. Several records in the definition and algorithm families read as authoritative definitions of technical terms; they are graph-generated prose, not verified reference material.
• Redundant column. voice_idx duplicates sys_idx on every record and carries no independent information in this file.
• Index referents are missing. As stated above, sys_idx/voice_idx are indices with no mapping table shipped in this file. Grouping by them is possible; interpreting them is not.
• The corpus is not deduplicated for semantic content. Task shapes repeat heavily by design (60,000 definition rows over a fixed node set), so near-duplicate wording recurs across records that differ only in the node they name.
• No train/validation/test split is provided. The file is a single flat split.
• This is the ungated version. The author's documentation describes a quality-gated derivative of the same generation run, omnilingua_train_v6_gated.jsonl, at 229,659 rows, described in 77_doctrine/THE_ORGANISM.md as follows, quoted verbatim: "(the OMNILINGUA quality gate scanned 315K rows and stripped 85K as padding, keeping 24.8% as CORE)". The "85K stripped" figure is consistent with the measured delta of 85,341 rows. The "24.8% as CORE" figure does not match this file: 229,659 is 72.91% of 315,000, and 229,659 would be 24.8% only of a ~926,044-row population, which does not correspond to this corpus. Treat the 24.8% figure as unexplained. The gated derivative is not included in this repository; only the ungated run is staged here.
• Provenance is documentation-derived. The generation description above comes from the author's doctrine files. The generator script itself was not located in the tree at staging time.
• Not determined from the content: what each sys_idx/voice_idx value denotes; whether the taxonomy's 243 subnodes map one-to-one onto any field; the exact generator version used for this file; and why 80 index values exist where the related chat corpus shows only 5 system prompts.
• Third-party material: a full-file scan found no reproduction of third-party copyrighted text. The corpus names third-party concepts as taxonomy nodes where a node name coincides with a general term (e.g. Existentialism, Nietzsche), but no quoted or excerpted source text was found — zero records contain copyright, ©, all rights reserved, or excerpt markers. Nothing was excluded on copyright grounds.

---

Files in this repository

File	Description
omnilingua_train_v6.jsonl	The corpus. 315,000 JSONL records, 172,775,995 bytes.
README.md	This dataset card.
LICENSE	CC BY 4.0.

---

Citation

@misc{betances_omnilingua_v6,
  author       = {Betances, Christopher},
  title        = {OmniLingua Training Corpus v6},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/CatQualia/omnilingua}},
  note         = {315,000 synthetic instruction/response records generated from a
                  hand-authored semantic taxonomy graph. CC BY 4.0.}
}

Plain text:

Betances, Christopher. OmniLingua Training Corpus v6. Hugging Face, 2026. https://huggingface.co/datasets/CatQualia/omnilingua

---

License

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). Copyright 2026 Christopher Betances (catqualia.com). See LICENSE.