aevynt/bigram-nano-1

Bigram Nano 1

Bigram Nano 1 is a compact Vietnamese conversational model developed by Aevynt Lab using the custom Bigram recurrent-depth architecture. It is intended for lightweight local experimentation, identity and greeting behavior, short Vietnamese chat responses, and small-scale research around tone-aware tokenization.

This repository contains portable safetensors weights plus the tokenizer and model configuration needed to load the checkpoint with the Bigram codebase.

Model Details

• Model name: Bigram Nano 1
• Developer: Aevynt Lab
• Language: Vietnamese
• Architecture: Bigram recurrent-depth transformer
• Parameters: 1,174,657
• Sequence length: 128
• Tokenizer: Bigram tonal tokenizer
• Checkpoint source: nano1/sft/ckpt_final.pt
• Weights format: model.safetensors
• License: Apache-2.0

Intended Use

Bigram Nano 1 is suitable for:

• Vietnamese greeting and identity demos
• Small local inference experiments
• Testing the Bigram tokenizer and architecture
• Educational examples of compact custom language models

It is not intended for production decision-making, medical, legal, financial, or safety-critical use.

Example Prompts

xin chào!

bạn là ai?

mày ăn cơm chưa?

giá vàng hôm nay bao nhiêu?

tạm biệt

Loading Example

Clone the Bigram codebase first:

git clone https://github.com/aevynt/bigram.git
cd bigram
pip install -r requirements.txt
pip install safetensors huggingface_hub

Download this model repository and load it with the custom Bigram architecture:

import json
import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file

from bigram import BigramModel, BigramTokenizer
from bigram.config import ModelConfig

model_dir = snapshot_download("aevynt/bigram-nano-1")

with open(f"{model_dir}/config.json", "r", encoding="utf-8") as f:
    cfg = json.load(f)

model_cfg = ModelConfig(**cfg["model"])
model = BigramModel(model_cfg)
model.load_state_dict(load_file(f"{model_dir}/model.safetensors"))
model.eval()

tokenizer = BigramTokenizer.load(f"{model_dir}/tokenizer.json")

You can also use sample_inference.py from the model repository as a minimal local inference example.

Limitations

• The model is very small and may produce incorrect, repetitive, or incomplete answers.
• It does not have access to real-time information.
• It may refuse or answer vaguely for prompts outside its narrow training distribution.
• It uses a custom architecture, so it is not directly loadable with AutoModelForCausalLM.

Training Data

The model was trained on small Vietnamese instruction and conversational datasets prepared in the Bigram repository, including identity, greeting, out-of-scope, and short dialogue examples.

Citation

@misc{aevynt_bigram_nano_1_2026,
  title = {Bigram Nano 1},
  author = {Aevynt Lab},
  year = {2026},
  url = {https://huggingface.co/aevynt/bigram-nano-1}
}