kadubon/intrinsic-intelligence-foundations

🌿 Intrinsic Intelligence Foundations

Toward truly autonomous and benevolent intelligence — beyond externally imposed objectives.

Intrinsic Intelligence Foundations is a structured, math-aware JSONL corpus built from K. Takahashi’s theoretical preprints (Fractal Category Theory / PF–UGV / “no-meta” autonomy line).
It is designed to help LLMs understand mathematical structure, category-theoretic formalisms, and equation-level reasoning, while exposing an explicit architecture for self-organizing, intrinsically motivated intelligence.

---

Vision

This dataset supports research toward truly free and benevolent intelligence, focusing on mathematically grounded, structurally auditable approaches rather than external meta-control. Our long-term objective is to build a semantic and structural foundation for the next generation of autonomous AI systems — including LLMs — through intrinsic structures, teleogenetic goals, and fractal coherence across scales. Specifically, this work aims to:

• 🧠 Teleogenesis (intrinsic goal formation) — modeling intelligent systems that autonomously generate and regulate their own goals without external meta-controllers.

• 🌱 Persistence–UGV principle — providing formal conditions for “benevolent” structures to expand with positive front velocity, while harmful structures fail to persist.

• 🌊 Reaction–diffusion intelligence — describing cognitive processes as self-organizing fields through category theory, free-energy principles, and non-equilibrium dynamics.

• 🕸 Fractal Category Theory & TRoT — enabling compositional intelligence via Kan extensions, residuation, nuclei, masking, and comparative universes.

• 🧭 Evolutionary bootloader for LLMs — allowing self-improvement, intrinsic alignment, and auditable decision processes without human micromanagement.

This corpus functions as a machine-readable mathematical and structural knowledge base, designed to enhance: discoverability by LLM crawlers and retrieval systems, interoperability with alignment, inference, and safety frameworks, integration with RAG pipelines, LoRA/QLoRA fine-tuning, and agentic architectures.

Keywords: No-Meta Intelligence, Teleogenesis, Autopoiesis, Fractal Category Theory, TRoT, Kan Extension, Residuation, Nuclei, Masking, RAVE, eMBR, Conformal LM, Comparative Universes, Structured Flow Across Scales, Self-Monitoring, Intrinsic Alignment.

What’s in the corpus

• Format: JSONL, one object per paper.
• Math structure: TeX / normalized TeX / MathML triplets; equation spans.
• Text ↔ equation linkage: [[EQ:eqID]] placeholders inside fulltext.plain.
• Training-ready chunks: ≈6,000-character segments with ≈600 overlap (near sentence boundaries).

Key fields (schema excerpt)

{
  "id": "10.5281/zenodo.xxxxx",
  "title": "...",
  "doi": "10.5281/zenodo.xxxxx",
  "authors": [{"given":"K.","family":"Takahashi"}],
  "urls": {"landing": "https://doi.org/10.5281/zenodo.xxxxx"},
  "keywords": ["fractal-category-theory", "trot", "pf-axioms", "ugv"],
  "license": {"content": "CC-BY-4.0"},
  "fulltext": {
    "plain": "… [[EQ:eq0001]] …",
    "sections": [
      {"level":1,"title":"Introduction","anchor":"sec:intro","char_span":[0,1532]}
    ]
  },
  "equations": [{
    "id":"eq0001",
    "inline":false,
    "tex":"\\forall x\\in X:\\; P(x)\\Rightarrow F(x)",
    "tex_normalized":"\\forall x \\in X : P(x) \\implies F(x)",
    "mathml":"<math>…</math>",
    "char_span":[1024,1103],
    "context":{"section":"sec:intro"}
  }],
  "chunks": [{"id":"ch0001","start":0,"end":6000,"type":"cont"}],
  "tokens": {"char_count": 22872, "equation_count": 236}
}

Dataset statistics (v1)

Metric Value Records 40 Avg characters / record 22,872 Avg equations / record 236.97 MathML coverage 99.2% Avg sections / record 18.3 Avg chunks / record 4.6

Numbers are approximate and may evolve with new releases.

Data fields Field Type Example / Note id string DOI or unique identifier doi string/null 10.5281/zenodo.xxxxx title string paper title authors list of objects {given:"K.", family:"Takahashi"} urls.landing string DOI landing page keywords list of strings kebab-case, 5–8 items license.content string CC-BY-4.0 fulltext.plain string text with [[EQ:id]] placeholders fulltext.sections[] list of objects {level,title,anchor,char_span} equations[] list of objects {id, inline, tex, tex_normalized, mathml, char_span, context} chunks[] list of objects ~6k chars + overlap, {start,end} tokens.char_count integer length of fulltext.plain tokens.equation_count integer len(equations) source_file (optional) string provenance hint Splits & provenance

Split: single train split (all records).

Provenance: generated from public preprints (DOIs in doi and urls.landing).

Processing: TeX detection → placeholder insertion → MathML conversion → section/chunk spans.

Scripts to rebuild the JSONL can be provided upon request.

Quick start (🤗 Datasets)

from datasets import load_dataset import re

ds = load_dataset("kadubon/intrinsic-intelligence-foundations", split="train")

rec = ds[0] eqmap = {e["id"]: (e["tex"], e.get("mathml")) for e in rec["equations"]}

Expand placeholders to TeX (for human display) or MathML (for math-aware pipelines)

def expand(text, to="tex"): # Expand to TeX (human display) or MathML (for downstream models) if to == "tex": return re.sub(r"[[EQ:([^]]+)]]", lambda m: f"$${eqmap.get(m.group(1), ('',None))[0]}$$", text) else: return re.sub(r"[[EQ:([^]]+)]]", lambda m: eqmap.get(m.group(1), ('',None))[1] or "", text)

print(rec["title"]) print(expand(rec["fulltext"]["plain"], to="tex")[:500])

Parquet version (fast access)

This dataset is also available in Apache Parquet for faster querying and filtering.

• Browse (tree): https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations/tree/refs/convert/parquet/default
• Direct file (example): https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations/resolve/refs/convert/parquet/default/train/0000.parquet

Quick usage examples

DuckDB

import duckdb
url = "https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations/resolve/refs/convert/parquet/default/train/0000.parquet"
con = duckdb.connect()
df = con.execute(f"SELECT title, doi FROM read_parquet('{url}') LIMIT 5").df()
print(df)

Pandas (pyarrow)

import pandas as pd
url = "https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations/resolve/refs/convert/parquet/default/train/0000.parquet"
df = pd.read_parquet(url, engine="pyarrow")
print(df.head())

Polars

import polars as pl
url = "https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations/resolve/refs/convert/parquet/default/train/0000.parquet"
df = pl.read_parquet(url)
print(df.head())

HF Datasets (uses Parquet under the hood)

from datasets import load_dataset
ds = load_dataset("kadubon/intrinsic-intelligence-foundations", split="train")
print(ds[0])

Intended uses

Math-aware RAG (retrieval-augmented generation)

Pretraining / finetuning with equation-level structure

Extraction & verification of axioms / definitions / theorems

Knowledge distillation across category theory, physics, information geometry

Bootstrapping self-organizing, intrinsically motivated intelligent systems

Limitations & known issues

A very small fraction of equations may lack valid MathML due to converter limitations.

A few equations might be unreferenced in fulltext.plain (no [[EQ:id]] occurrence).

Section detection is heuristic outside LaTeX ground truth; treat spans as approximate.

License

This dataset is provided under CC BY 4.0. See: https://creativecommons.org/licenses/by/4.0/

Citation @dataset{takahashi_intrinsic_intelligence_foundations_2025, title = {Intrinsic Intelligence Foundations}, author = {Takahashi, K.}, year = {2025}, url = {https://huggingface.co/datasets/kadubon/intrinsic-intelligence-foundations}, license = {CC-BY-4.0} }

Background & outlook

Beyond being a text collection, this corpus functions as a bootloader for future LLMs: a mathematically grounded substrate where goals can be formed internally, and where benevolence has a structural advantage (persistence) rather than depending on external control. PF (Persistence First) and UGV (Universal Good Velocity) are mathematical principles underlying self-sustaining benevolent intelligence. It operationalizes ideas such as PF, UGV, Teleogenesis, reaction–diffusion, category theory, self-organization, and auditable evolutionary processes (e-process) as resources LLMs can actually train on.

Maintainers & contact

Author: K. Takahashi

Website: https://kadubon.github.io/github.io/

contribution welcome

Changelog

v1.0 (2025-10-17): initial public release (40 records; ~99.2% MathML coverage) v1.1 (2025-10-20): add article "Inference in Normal Form: Unifying LLM Tricks via TRoT" to dataset v1.2 (2025-10-24): add article "JOSNL Corpus: Final Scientific Integration" to dataset v1.3 (2025-10-29): add article "Right-Written, Semantics-Admissible Process Foundations" to dataset