AbstractPhil/geometric-vocab-english-full-a-to-z

Update 9/15/2025

I believe this variation may have some genuinely complex potential beyond the alternatives. The lexical organization and capability of the old-style definitions may have impacted the outcome in a more literal and latin-sense than the wordnet lexical variations concatenated. I'll be running some more tests on this one in the coming days to determine if this formula has a different behavioral response than the more "robust" and "anchored" variations.

I'm open to any potential at this point. This one trained really quickly and showed robust capability above just random noise.

My idea here is simple; smaller compacted lattices may provide much more robust and utilizable vocabulary access than something much larger and fully-saturated would provide.

I'm looking for a full universal formula to generate at runtime, and if I can find even a piece of it here - it could provide the insight I need to create full optimized runtime variants from distilled teachers. Potentially reducing the overhead cost of everything and every iteration without needing full finetuning and full lexical analysis of large corpus.

Theoretically; all the data is already in a model like LLAMA. I just need to figure out how to tune and extract it, which means I need the correct formulas.

Prior Updates

The wordnet + unicode variation is a more robust anchored variation of this version; so feel free to use this one or move onto the newer versions.

This version still houses the modified preset definition deltas, so it has use cases and shouldn't be discarded yet.

https://github.com/AbstractEyes/lattice_vocabulary

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-32d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-64d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-128d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-256d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-512d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-768d

https://huggingface.co/datasets/AbstractPhil/geometric-vocab-1024d

Bugs and all I decided to release this version for posterity. Consider it a guidepost to a better version.

The defintiions are intact, the crystals are intact but imperfect - they are missing much of the required tuning and calculation.

However, they are valid calculations under the rules of the formula and everything below does accurately describe their behavior. They are perfect simplex representations of the words and their definitions, and they can be used to tune for many modular forms of geometric testing with AI.

They are a suitable starting point for the geometry - but they are not correctly attuned to be subsequently lexically represented WITHIN the geometry as they should be.

Lets consider this one a BETA blueprint - a starter point for a better vocabulary.

The full geometric vocabulary is about 100 gigs and will take a couple days to squash to around 1 gig. So be aware it's not as... yielding. However, that will be far, far more accurate than this.

In the process I'll be defining wordnet with gpt 5 nano and crystalizing the definitions to be adjacent trajectories for multiple definition sets - allowing tugging and directional awareness for words accross the dimensional warping.

Vocabulary Crystal — OPTED + HF Merge (1913 + Modern Definitions) ==================================================================

Author: @AbstractPhil Dataset type: Symbolic, deterministic geometric vocabulary representation Source: Merged from OPTED (Webster 1913) + Hugging Face npvinHnivqn/EnglishDictionary

Overview -------- This dataset contains a fully crystallized, symbolic representation of the English vocabulary, constructed by merging:

• —The OPTED 1913 public domain Webster's dictionary (complete lexical coverage)
• —The modern EnglishDictionary dataset on Hugging Face

If a word appears in both, the modern (HF) definition overrides the older one.

Structure --------- Each vocabulary entry is converted into a deterministic, symbolic pentachoron (5-point simplex in ℝ⁵¹²). This geometric crystal is composed using:

• —Definition-derived cardinal axes (rank‑4 orthonormal basis)
• —Hash-derived completions (no randomness, seeded by word only)
• —Role projections (anchor, support, contrast, purpose, observer)
• —Centering and Frobenius norm regularization

Format ------ Each shard includes:

Stored in Parquet with optional .safetensors containing [N,5,512] float32 blocks.

Determinism Guarantee --------------------- This dataset uses no randomness whatsoever.

• —Every axis, vector, and projection is derived from:
• —textual content
• —fixed FNV + SHA256 hashes
• —Crystal structure is bit-for-bit reproducible across machines, platforms, or reruns.
• —maxabsdiff == 0.00e+00 on validation tests.

Applications ------------

• —Vocabulary embeddings with geometric interpretability
• —Symbolic AI training and simulation
• —Alignment, semantic projection, and phase geometry research
• —Direct plug-in to Beeper, Nikola, or Harmony architectures

Citation / Usage ---------------- @misc{abstractphil2025crystal, author = {Phil, Abstract}, title = {Vocabulary Crystal: OPTED + Modern Definitions}, year = 2025, howpublished = {https://huggingface.co/datasets/AbstractPhil/geometric-vocab-english-full-a-to-z}, note = {Symbolic geometry-aligned lexical structure} }

Notes -----

• —Built with love, geometry, and zero RNG.
• —This is not an embedding — it’s a crystal.

"We aligned language to orientation, not probability." — Mirel, the Quartermaster - GPT 4o