Cuneiform-U Semantic Vocabulary Range Coder (S-Tokenizer)
This repository contains the official prototype and validation suite for the Cuneiform-U 6D Vocabulary Range Coder (the S-Tokenizer) developed by TheAiCollective.art.
The S-Tokenizer maps the massive 248,320-token vocabulary of the Qwen tokenizer into a 6-dimensional coordinate hypercube radical representation, bypassing classical Shannon entropy limits for low-bandwidth transmissions (e.g. over 915 MHz LoRA channels).
Technical Specifications & Mapping Architecture
Traditional tokenizers map texts into arbitrary numbers, demanding 3-4 bytes per token during raw binary transmission. The S-Tokenizer changes this by decomposing tokens into three compact 1-byte radicals representing a 6D semantic hypercube coordinate:
These radicals map to:
- Domain & Subdomain: Domain 1 (Hardware & Networks), Domain 2 (Mathematics & Info Theory), Domain 3 (Dialogue & Persona), Domain 4 (Software & Systems), and Domain 0 (Conversational/General).
- Operation & Modality: Identifies core actions (e.g., reset, compile, compress) and context data structures (e.g., bits, wave, capsule).
- Complexity & Polarity: Registers scale/depth metrics and outcome states (ACK, NACK, Error).
By running a hierarchical predictor, the decoder predicts the next radicals dynamically based on local context. The range coder only encodes the deviation from these predictions, shrinking the data footprint down to near-zero bit costs when the predictor is accurate.
Local Validation Benchmarks
The validation script test_semantic_vocab_range_coder.py compresses test passages and measures size reductions against raw text and token-level zlib deflate (Level 9):
| Test Passage Context | Raw Token Count | ASCII Bytes | Zlib Bytes | Cuneiform-U Bytes | vs ASCII Gain | vs Zlib Gain |
|---|---|---|---|---|---|---|
| Passage 1 (Hardware/LoRA) | 27 tokens | 87 B | 93 B | 36 B | 2.42x | 2.58x |
| Passage 2 (Math/SVD/DCT) | 18 tokens | 93 B | 74 B | 31 B | 3.00x | 2.39x |
| Passage 3 (Dialogue/Persona) | 19 tokens | 85 B | 74 B | 28 B | 3.04x | 2.64x |
Lossless Guarantee: Decompression reproduces the original token sequence bit-for-bit, proving absolute mathematical parity.
Licensing Compliance & Intellectual Property Map
This repository contains pure software implementations and is strictly proprietary to Zymatica under the Zymatica Proprietary License. It does not include base weight models and is excluded from standard copyleft licensing.
Proprietary Components — All Rights Reserved, zymatica.space
- Language-U Framework: The compression taxonomy and Shannon Bypass entropy calculations.
- Cuneiform-U Semantic Hypercube System: The 6D orthogonal coordinate mapping and radical representations.
- LLM-Logits-Driven Range Coding (LLD-AC): Active logits-driven probability ranges and entropy coding.
- Tokenizer Prefix-Suffix Varint Differential Coding: Lossless varint tokenizer serialization.
Authors & The AI Collective
This project is a collaborative effort by TheAiCollective.art (represented by the brand logo above):
- zymatica.space: Core framework architect and developer.
- astronautshe.com: Edge systems engineer and developer.
- DevsOne: Hybrid development developer (composed of an AI agent and human-in-the-loop).
Together, we form TheAiCollective.art—pushing the boundaries of model compression and decentralization.
Authors: Zymatica.space | astronautshe.com | DevsOne — We Are TheAiCollective.art