bvv241-abs: Unified Unicode Tokenizer (SOTA Intersection) with Frozen Embeddings and Extended Vector Dim (4096)

Tokenizer Description

This tokenizer is based on a hybrid vocabulary:

This tokenizer uses a strictly structured Unicode mapping scheme:

• Plane 0 (0–65535): All single Unicode code points (monograms) are mapped 1:1 to token codes, directly matching standard Unicode BMP.
• Private and unused code ranges (Plane 0 high + supplementary, e.g., 0xE000–0xF8FF and 65536–131071):
  • All multi-character tokens (bigrams, trigrams, SOTA model token strings) are placed exclusively in these ranges.
• This design achieves total, lossless Unicode text coverage, with all multi-symbol tokens isolated above the core Unicode range.
• Tokenizer created from the intersection of token text across leading SOTA models
• Includes o200k_base, cl100k_base, Mistral-Nemo, QwQ-32B, DeepSeek-R1, Qwen3-32B vocabularies,
• Vocabulary size: 131,072 tokens,
• Embedding dimension: 4096.

The associated normalized_embeddings_weights.pt file contains a [vocab_size x embed_dim] matrix of precomputed, L2-normalized, frozen embeddings.
No semantic information is encoded; embeddings remain fixed throughout LM pretraining. No training or adaptation; suitable for plug-and-play use in research on embedding-free semantic emergence and modular LMs.

How to Get Started with the Tokenizer

from transformers import AutoTokenizer

from huggingface_hub import hf_hub_download

import torch

tokenizer = AutoTokenizer.from_pretrained('Bochkov/bvv241-abs')


emb_path = hf_hub_download(
    repo_id="Bochkov/bvv241-abs",
    filename="normalized_embeddings_weights.pt"
)

embeddings = torch.load(emb_path)

🧑‍🔬 Citation & Concept

If you use this model or the underlying concepts in your research, please cite our work:

@misc{bochkov2025emergentsemanticstokenembeddings,
      title={Emergent Semantics Beyond Token Embeddings: Transformer LMs with Frozen Visual Unicode Representations}, 
      author={A. Bochkov},
      year={2025},
      eprint={2507.04886},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.04886}, 
}

This work demonstrates that transformer blocks, not token embeddings, carry the semantic burden in LLMs — a step toward modular, fusable, multilingual LMs.