emergent-semantics-model-uni-glyph-335m

Emergent Semantics — Model_UNI_GLYPH (335M)

This repository provides Model_UNI_GLYPH (335M) — a decoder-only Transformer language model where the entire input embedding layer is frozen and initialized from visual Unicode glyph representations (rendered glyph bitmaps → PCA projection → L2 normalization).
The model is released as part of 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) -

📚 Blog Article

Primary goal: research and ablations on where semantic structure emerges in Transformer LMs when input embeddings are non-trainable and non-semantic.

---

Key idea

Standard LMs learn token embeddings jointly with the rest of the network. In this work, embeddings are treated as fixed structural primitives rather than “meaning vectors”.

Model_UNI_GLYPH uses:

• Frozen nn.Embedding weights (no gradient updates)
• Embeddings derived from Unicode glyph images (deterministic, precomputed)
• A standard decoder-only Transformer backbone trained normally

This isolates semantics as an emergent property of Transformer layers, rather than a property of learned input embeddings.

---

Model summary

• Architecture: decoder-only Transformer (GPT-like)
• Hidden size (d_model): 1024
• Layers: 16
• Heads: 32
• Positional encoding: rotary embeddings
• Activation: GELU
• Input embeddings: frozen, glyph-based (visual Unicode, PCA-projected)
• Output head: not tied to the input embeddings (trained separately)
• Vocabulary size: 65,536
• Tokenizer: Bochkov/bvv241-2-3

---

Intended use

This model is intended for:

• Research on emergent semantics in Transformers
• Studying optimization effects of frozen vs trainable embeddings
• Multilingual / Unicode-heavy text experiments where deterministic coverage is useful
• Reproducible ablations on embedding initialization

Not intended for production deployments. It is a research artifact trained under constrained compute/data to enable controlled comparisons.

---

Evaluation (reported)

These are the key numbers reported for this checkpoint:

• MMLU: 23.81 (5-shot)
• CommonsenseQA (C-SENSE): 19.75
• ARC-E: 22.5
• ARC-C: 22.1

Note: The objective of the project is not SOTA performance. The primary comparison is against architecturally identical baselines differing only in whether embeddings are trainable.

---

How to use (Transformers)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Bochkov/emergent-semantics-model-uni-glyph-335m")
model = AutoModelForCausalLM.from_pretrained("Bochkov/emergent-semantics-model-uni-glyph-335m", 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:Tokyo (city)

---

Training overview (high level)

• Training data: multilingual Wikipedia subsets + a small portion of SFT-style QA data (see paper for details)
• Scale: ~4B tokens (resource-constrained setting for controlled comparisons)
• Hardware: H100 80GB (reported setup)

---

Limitations

• Research-focused; not tuned for safety, factuality, or instruction following.
• Benchmark scores reflect a constrained training regime; do not compare directly to large-scale LLMs trained on hundreds of billions of tokens.
• Frozen embeddings may encode surface/form biases (e.g., length/density effects from glyph rasterization + PCA).

---

Related repositories

• Paper model collection:
  https://huggingface.co/collections/Bochkov/emergent-semantics-beyond-token-embeddings
• Tokenizer:
  https://huggingface.co/Bochkov/bvv241-2-3
• 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}, 
}