README.md

---
license: apache-2.0
tags:
  - text-generation
  - causal-lm
  - transformer
  - research
  - interpretability
  - multilingual
  - unicode
  - frozen-embeddings
  - ablation
language:
  - multilingual
library_name: transformers
pipeline_tag: text-generation
---

# Emergent Semantics — Model_16_FLOAT (269M)

This repository provides **Model_16_FLOAT (269M)** — an **ablation model** from the paper:

[📚 Paper (Emergent Semantics Beyond Token Embeddings: Transformer LMs with Frozen Visual Unicode Representations)](https://huggingface.co/papers/2507.04886) -

[📚 Paper (Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate)](https://huggingface.co/papers/2507.07129) -

[📚 Blog Article](https://huggingface.co/blog/Bochkov/emergent-semantics-beyond-token-embeddings)

This checkpoint is designed to study the effect of **normalization / PCA-style processing** in a *minimal* frozen embedding setting.

Unlike **Model_UNI_GLYPH**, this model does **not** use glyph-based embeddings. Instead, it uses a **frozen 16-dimensional float embedding** per token.

---

## Key idea (what this ablation tests)

This model isolates the impact of having **float** frozen embeddings (with **PCA + normalization**) versus the strictly **binary token-ID** variant (**Model_16_BIT**):

- **`n_embed = 16`** per token (**float components**, not binary)
- Embedding vectors are **precomputed** (PCA + L2 normalization) and then **frozen**
- The embedding layer is never updated (`requires_grad=False`)
- 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`

This lets you test whether the model’s behavior changes when the frozen token “identifier” is:
- discrete + purely ID-like (**16-bit**), vs
- continuous + normalized (**16-float**)

---

## 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 primarily 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 **Transformer backbone is the same** (layers/heads/d_model), but the embedding table is much smaller, reducing total parameters.

---

## 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**, `n_embed=16` (**float**, PCA + L2 normalized), expanded to 1024 by repetition (non-trainable)
- **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)

```python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Bochkov/emergent-semantics-model-16-float-269m")
model = AutoModelForCausalLM.from_pretrained("Bochkov/emergent-semantics-model-16-float-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:A temperature in

```

---

## Intended use

Research only, especially for:

- Comparing **Model_16_FLOAT** vs **Model_16_BIT** (effect of continuous normalized vectors vs binary ID)
- Comparing **Model_16_FLOAT** vs **Model_UNI_GLYPH** (effect of glyph-derived structure vs minimal vectors)
- Studying emergent semantics when embeddings are **frozen and non-semantic**

Not intended for production deployment.

---

## Related links

- **Model collection (paper artifacts):**  
  https://huggingface.co/collections/Bochkov/emergent-semantics-beyond-token-embeddings
- **UNI_GLYPH main model:**  
  https://huggingface.co/Bochkov/emergent-semantics-model-uni-glyph-335m
- **16-bit ablation:**  
  https://huggingface.co/Bochkov/emergent-semantics-model-16-bit-269m
- **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}, 
}
```