Meeting Minutes Generator

This is an example project that takes a meeting transcript and generates a summary (meeting minutes). It utilizes a local model from Hugging Face with the transformers dependency. The selected model is Deepseek R1, running locally on a CPU. The input transcript is tokenized before processing.

Installation & Setup

1. Prepare the Environment

1. Create a .env file in the project root with the following content:

   HF_TOKEN=your-hugging-face-token
   TRANSCRIPT_FILE_NAME=transcript_example.txt

2. Place the transcript file (transcript_example.txt) inside the assets folder.

2. Setup Virtual Environment

Run the following commands:

python3 -m venv venv
source venv/bin/activate  # On macOS/Linux
venv\Scripts\activate  # On Windows

3. Install Dependencies

pip3 install -r requirements.txt

4. Run the Application

python3 main.py

The generated meeting minutes will be printed on the screen.

Screenshot

Code Design

To structure the code efficiently, a pipeline pattern was quickly implemented. This allows better organization and modularity, making it easier to extend or modify in the future.

Running on GPU (Optional)

If executing on a GPU, it is recommended to apply quantization for optimized performance. An example configuration using BitsAndBytesConfig is provided below:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,  # Double quantization (32b -> 8b -> 4b)
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4"  # Normalized Floating Point 4-bit
)

model = AutoModelForCausalLM.from_pretrained(
    DEEP_SEEK_R1_DISTILL,
    trust_remote_code=True,
    device_map="auto",
    quantization_config=quant_config,
)

This reduces memory usage and speeds up inference while maintaining accuracy.