metadata

title: Refsheet Chat
emoji: 💬
colorFrom: gray
colorTo: green
sdk: gradio
sdk_version: 5.21.0
app_file: app.py
pinned: false
license: mit
short_description: Chat with a character via reference sheet!

RefSheet Chat -- Chat with a character via reference sheet

Upload a reference sheet of a character, RefSheet Chat will try to understand the character through the reference sheet, and talk to you as that character. RefSheet Chat can run locally to ensure privacy.

Website: https://refsheet.chat

Tutorial slide (in Chinese) can be found in https://snowkylin.github.io/talks/

RefSheet Chat is a demo of Gemma 3, demonstrating its excellent vision and multilingual capability.

Environment Configuration

Submit a Gemma Access Request from https://huggingface.co/google/gemma-3-4b-it. The access should be granted immediately with an email notification. After that, the model page will show

Gated model: You have been granted access to this model

Create conda environment with pip and Python 3.12

conda create -n transformers_gemma pip python=3.12
conda activate transformers_gemma

Install HuggingFace Transformers for Gemma 3:

pip install git+https://github.com/huggingface/[email protected]

Install PyTorch

On Nvidia GPU (with CUDA 12.6):

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

On CPU:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Create an User Access Token from https://huggingface.co/docs/hub/security-tokens, then log in to your HuggingFace account with huggingface-cli:

huggingface-cli login

Copy-paste your access token and press enter.

Packing

See https://github.com/whitphx/gradio-pyinstaller-example for more details

Create a hook file runtime_hook.py including environment variables

# This is the hook patching the `multiprocessing.freeze_support` function,
# which we must import before calling `multiprocessing.freeze_support`.
import PyInstaller.hooks.rthooks.pyi_rth_multiprocessing  # noqa: F401
import os

if __name__ == "__main__":
    os.environ['PYINSTALLER'] = "1"
    os.environ['HF_ENDPOINT'] = "https://hf-mirror.com" # optional, HF mirror site in China
    os.environ['HF_TOKEN'] = "hf_XXXX"  # HF token that allow access to Gemma 3
    # This is necessary to prevent an infinite app launch loop.
    import multiprocessing
    multiprocessing.freeze_support()

Then

pyi-makespec --collect-data=gradio_client --collect-data=gradio --collect-data=safehttpx --collect-data=groovy --runtime-hook=./runtime_hook.py app.py

open app.spec and add

a = Analysis(
    ...,
    module_collection_mode={
        'gradio': 'py',  # Collect gradio package as source .py files
    }
}

then pack the environment

pyinstaller --clean app.spec

finally copy the win32ctypes folder from your conda environment

C:\Users\[Your-User-Name]\miniconda3\envs\[Your-Env-Name]\Lib\site-packages

to dist/app/_internal.

Run app.exe in dist/app and it should work.