Bochkov/emergent-semantics-model-64-float-272m

Emergent Semantics — Model_64_FLOAT (272M)

This repository provides Model_64_FLOAT (272M) — 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) -

This checkpoint tests whether language modeling and semantic structure can emerge when the entire input embedding layer is frozen and contains no semantic or glyph/visual information.

Compared to Model_64_BIT, this model uses the same embedding dimensionality (n_embed=64) and the same “unique per token” construction, but the embedding vectors are floating-point (after a deterministic projection/normalization step), rather than raw binary components.

---

Key idea (what this ablation tests)

• Each token is assigned a frozen 64-dimensional float vector (n_embed=64).
• The vectors originate from random per-token patterns and are constructed to guarantee a unique ID per token (no collisions by design).
• A deterministic post-processing step (e.g., PCA/projection + normalization) converts the raw patterns into float embeddings and standardizes their scale.
• The embedding layer is frozen throughout training (requires_grad = False).

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

This keeps the Transformer backbone identical while isolating the role of embedding trainability and embedding content.

---

Important: parameter count difference (vs 335M models)

This checkpoint has ~272M parameters, while models with a standard n_embed=1024 embedding table (e.g. UNI_GLYPH / unfrozen baselines) are ~335M.

The reduction is primarily due to the smaller embedding matrix:

• Standard embedding params: vocab_size * 1024 = 65536 * 1024 ≈ 67.1M
• This model’s embedding params: vocab_size * 64 = 65536 * 64 ≈ 4.19M

So the Transformer backbone is the same, but the embedding table is much smaller, lowering 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, float, n_embed=64, expanded to 1024 by repetition (non-trainable)
• Embedding initialization: random per-token patterns → deterministic projection/normalization → float vectors (unique per token, no collisions)
• Output head: not tied to the input embeddings (trained separately)

---

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-64-float-272m")
model = AutoModelForCausalLM.from_pretrained("Bochkov/emergent-semantics-model-64-float-272m", 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:Japan
#    </s><|

---

Intended use

This model is intended for research only, especially for:

• Comparisons vs Model_UNI_GLYPH (glyph/PCA frozen embeddings) and vs trainable-embedding baselines
• Ablations comparing binary vs float frozen identifier embeddings at the same n_embed
• Studying whether semantic structure emerges in Transformer blocks when the input embedding space is a random-but-unique float code

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}, 
}