Bochkov / emergent-semantics-model-16-bit-269m

Emergent Semantics — Model_16_BIT (269M)

This repository provides Model_16_BIT (269M) — an ablation model from the paper:

📚 Paper (Emergent Semantics Beyond Token Embeddings: Transformer LMs with Frozen Visual Unicode Representations) -

📚 Paper (Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate) -

The purpose of this checkpoint is to isolate whether trainable/semantic input embeddings are necessary at all .

Unlike Model_UNI_GLYPH , this model uses a frozen 16-dimensional binary embedding that contains no glyph / visual information — it effectively encodes only the token identity .

Key idea (what this ablation tests)

• Vocabulary size is 65,536 , so a token index fits exactly into 16 bits .
• Each token is mapped to a fixed 16-bit vector (binary components), i.e. a deterministic token-ID representation.
• This embedding layer is frozen for the entire training run.

To match the Transformer hidden size, the 16-dim embedding is expanded to 1024 via a non-trainable repetition : repeat_interleave(64) → 16 * 64 = 1024 .

So the model receives a full d_model=1024 input vector, but the learned semantic geometry is not coming from the embedding table .

Important: parameter count difference (vs 335M models)

This checkpoint has ~269M parameters , while models with a standard n_embed=1024 embedding table (e.g. UNI_GLYPH / unfrozen baselines ) are ~335M .

This difference is expected and comes almost entirely from the embedding matrix size:

• Standard embedding params: vocab_size * 1024 = 65536 * 1024 ≈ 67.1M
• This model’s embedding params: vocab_size * 16 = 65536 * 16 ≈ 1.0M

So the model is architecturally identical in the Transformer backbone (layers/heads/d_model), but has a much smaller frozen embedding table , which reduces total parameter count.

Model summary

• Architecture: decoder-only Transformer (GPT-like)
• Hidden size ( d_model ): 1024
• Layers: 16
• Heads: 32
• Positional encoding: rotary embeddings
• Activation: GELU
• Tokenizer / vocab size: 65,536 (bvv241-2-3 compatible)
• Input embeddings: frozen , binary , n_embed=16 , expanded to 1024 by repetition (non-trainable)
• Output head: not tied to the input embeddings (trained separately)

Files in this repo (embedding reference)

To make this ablation fully inspectable/reproducible, the explicit frozen embedding values are provided:

• embeddings.txt (human-readable reference):
  https://huggingface.co/Bochkov/emergent-semantics-model-16-bit-269m/raw/main/embeddings.txt

This is convenient for verifying the exact token→vector mapping for each token ID.

Note: Embeddings are shipped in this model repo (even though the tokenizer exists as a separate HF repo) to keep the model+embedding mapping self-contained and unambiguous.

Tokenizer

The intended tokenizer is bvv241-2-3 (same vocab size and indexing):

• https://huggingface.co/Bochkov/bvv241-2-3

You may load the tokenizer either from this model repo (if included) or from the standalone tokenizer repo. The key requirement is exact vocab alignment .

How to use (Transformers)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Bochkov/emergent-semantics-model-16-bit-269m")
model = AutoModelForCausalLM.from_pretrained("Bochkov/emergent-semantics-model-16-bit-269m", trust_remote_code=True).to('cuda')

inputs = torch.tensor([tokenizer.encode("Question: What is the capital of Japan?\nAnswer:")], dtype=torch.long, device='cuda')

outputs = model.generate(
    inputs, 
    max_new_tokens=10,
    do_sample=False
)
print(tokenizer.decode(outputs[0].tolist()))

#Question: What is the capital of Japan?
#Answer:Nagano Prefecture

Intended use

This model is intended for research only , especially for:

• Controlled comparisons vs Model_UNI_GLYPH (glyph/PCA frozen embeddings) and vs trainable-embedding baselines
• Studying whether semantic structure emerges in Transformer blocks even when the input embedding layer is purely structural / token-ID-like
• Interpretable ablation experiments on embedding initialization and freezing

Not intended for production deployment (no instruction tuning, safety tuning, or factuality guarantees).

Related links

• Model collection (paper artifacts):
  https://huggingface.co/collections/Bochkov/emergent-semantics-beyond-token-embeddings
• UNI_GLYPH main model (frozen visual glyph embeddings):
  https://huggingface.co/Bochkov/emergent-semantics-model-uni-glyph-335m
• Tokenizer collection:
  https://huggingface.co/collections/Bochkov/tokenizers
• Code (GitHub):
  https://github.com/AVBochkov/Embeddings

🧑‍🔬 Citation & Concept

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

@article{
      bochkov2025emergent,
      title={Emergent Semantics Beyond Token Embeddings: Transformer {LM}s with Frozen Visual Unicode Representations},
      author={Andrey Bochkov},
      journal={Transactions on Machine Learning Research},
      issn={2835-8856},
      year={2025},
      url={https://openreview.net/forum?id=Odh8IynO1o},
      note={}
}
@misc{bochkov2025growingtransformersmodularcomposition,
      title={Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate}, 
      author={A. Bochkov},
      year={2025},
      eprint={2507.07129},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2507.07129}, 
}