nemo_bvv_zh

This is nemo_bvv_zh, a Chinese language GPT-style model trained with fully precomputed and frozen token embeddings using the Mistral/Nemo tokenizer (visual appearance-based). Specifically designed to demonstrate the compatibility of SOTA tokenizers in the fixed-embedding paradigm.

Model Details

Parameters: ~400M
Vocabulary: 131072 tokens (Mistral Nemo, precomputed embeddings)
Frozen Embeddings: Yes (not fine-tuned or trained)
Architecture: 12 layers, 12-head, hidden size 1024
Corpus: Small-scale, 10% SFT
Purpose: Research, foundation for MoE fusion

Performance

MMLU (average): 6.62%
ARC-e: 23.53%
ARC-c: 24.48%
Commonsense-QA: 19.27%
SQUAD: 6.05%
BLEU [en-zh]: 0.87% / [zh-en]: 4.05%

Limitations

Trained on a tiny research corpus for demonstration only.
Accuracy is far below large-scale production models.
Not suitable for commercial or mission-critical tasks.
No safety or fairness tuning.

🧑‍🔬 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}, 
}

@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}, 
}

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

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained('Bochkov/nemo_bvv_zh', trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained('Bochkov/nemo_bvv_zh', trust_remote_code=True).to('cuda')
inputs = tokenizer("你好，世界！ ", return_tensors="pt").to('cuda')
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Bochkov
/

nemo_bvv_zh