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.
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.
