vi-humanizer · Gemma 4 E2B

A small assistant whose Vietnamese reads as if a native speaker wrote it. Ask it anything (explain, advise, plan, write an email or a post, translate), or give it an AI-written draft to rewrite. Answers come with the right tone and xưng hô, without chatbot preambles, closing offers or translationese.

About 2B effective parameters, built on Gemma 4 E2B. Weights in bf16 for transformers and vLLM, and GGUF files for llama.cpp, LM Studio and Ollama, down to 3.4 GB.

Quick start

Local app (offline): ContextBoxAI/vi-humanizer is a web app with a rewrite view (draft and rewrite side by side, with the changes and a numbers check) and a chat view.

git clone https://github.com/ContextBoxAI/vi-humanizer && cd vi-humanizer
./install.sh          # macOS, Linux; .\install.ps1 on Windows

It installs llama.cpp if needed, downloads the Q6_K file once and opens the app in your browser.

llama.cpp:

llama-server -hf contextboxai/vi-humanizer-gemma-4-e2b:Q6_K --jinja -ngl 99    # OpenAI-compatible API on :8080

The GGUF files carry their own chat template and sampling defaults (temperature 1.0, top-p 0.95, top-k 64), so no other flags are needed. LM Studio picks them up too. For Ollama, use the Modelfile from the app repo.

vLLM: vllm serve contextboxai/vi-humanizer-gemma-4-e2b

transformers:

from transformers import AutoTokenizer, Gemma4ForConditionalGeneration
tok = AutoTokenizer.from_pretrained("contextboxai/vi-humanizer-gemma-4-e2b")
model = Gemma4ForConditionalGeneration.from_pretrained("contextboxai/vi-humanizer-gemma-4-e2b", dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "Viết giúp mình tin nhắn xin sếp nghỉ chiều nay để đưa con đi khám."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=512, do_sample=True)[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

No system prompt is needed; one you add (an operator role, a persona) is respected. Thinking mode is off in the chat template.

Files

License

A derivative of Gemma, subject to the Gemma Terms of Use.