FragenAntwortLLMGPU

FragenAntwortLLMGPU is a Python package for processing PDF documents and generating efficient Question & Answer (Q&A) pairs using LLM backends on CPU or GPU. The generated Q&A pairs can be used for LLM fine-tuning.

It supports:

• Mistral (GGUF via CTransformers) — local GGUF models
• Qwen (Hugging Face Transformers) — HF models like Qwen2.5 Instruct

---

Table of Contents

• Installation
• Usage
• Model selection
• Features
• Contributing
• License
• Authors

---

Installation

Install the package

pip install FragenAntwortLLMGPU

Notes on PyTorch / GPU

This project uses PyTorch. Install a PyTorch build that matches your system (CPU or your CUDA version).
If you already have PyTorch installed, you can skip this.

PyTorch install guide:

• https://pytorch.org/get-started/locally/

Optional backend dependencies

Qwen (Transformers backend)

pip install transformers accelerate

Mistral GGUF (CTransformers backend)

pip install ctransformers

---

Usage

Python example

from FragenAntwortLLMGPU import DocumentProcessor

processor = DocumentProcessor(
    book_path="/path/to/your/book/",      # directory containing the PDF
    temp_folder="/path/to/temp/folder",
    output_file="/path/to/output/QA.jsonl",
    book_name="example.pdf",
    start_page=9,
    end_page=77,
    gpu_layers=100,
    number_Q_A="five",  # written number: "one", "two", ...
    target_information="foods and locations",
    max_new_tokens=1000,
    temperature=0.1,
    context_length=2100,
    max_tokens_chunk=800,
    arbitrary_prompt="",
    model="mistral",  # default
)

processor.process_book()
processor.generate_prompts()
processor.save_to_jsonl()

---

Model selection

Default (Mistral GGUF via CTransformers)

The default backend is GGUF via CTransformers.

Example model source (GGUF):

• https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-GGUF

Use:

processor = DocumentProcessor(..., model="mistral")

Qwen (Hugging Face Transformers)

Use:

processor = DocumentProcessor(
    ...,
    model="qwen",
    # optionally override the HF model id
    hf_model_id="Qwen/Qwen2.5-7B-Instruct",
)

---

Features

• Extracts text from PDF documents
• Splits text into manageable chunks for processing
• Generates efficient Q&A pairs based on specific target information
• Supports custom prompts for question generation
• Runs on CPU and GPU (depending on backend and installation)
• Multilingual input: accepts PDF books in French, German, or English and generates Q&A pairs in English

---

Contributing

Contributions are welcome! Please fork the repository and submit pull requests.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Authors

Mehrdad Almasi, Lars Wieneke, and Demival Vasques Filho