Text Summarizer Using Phi 3

This repository contains a text summarizer that leverages the Phi 3 model to generate summaries from textual artifacts. The summarizer can process large texts, generate summaries, and convert them into markdown and PDF formats.

$~$

Table of Contents

• Features
• Installation
• Usage
  • General Runner Usage
  • Example of Usage
• Configuration
• Input Data Format
• Output
• Contributing

$~$

Features

• Multi-GPU Support: Efficiently utilize multiple GPUs for faster processing.
• Customizable Commands: Easily modify default commands and arguments.
• Markdown Generation: Convert JSON summaries into well-structured Markdown documents.
• PDF Conversion: Transform Markdown files into PDF documents.
• Organized Summaries: Automatically organize summaries by author and generate a table of contents.

$~$

Installation

1. Fork the repository.

git clone https://github.com/your-username/TextSummarizer.git
cd TextSummarizer

2. Install the required dependencies:

pip install torch transformers nltk tqdm

3. Download the NLTK punkt tokenizer:

import nltk
nltk.download('punkt')

4. install CUDA toolkit and appropriate GPU drivers for your system to enable GPU acceleration.

5. Ensure you have sufficient disk space and RAM to load and run the Phi-3 model.

6. Run the main script llm_general_multi_process_gpu.py with the appropriate command-line arguments as shown in the example usage.

$~$

Usage

General Runner Usage

To run the summarizer, use the following command:

python3 utility/llm_general_multi_process_gpu.py \
    --input_json ./youtube/videos.json \
    --output_dir ./youtube/partitions \
    --gpu_devices 5 6 7 \
    --processes_per_gpu 2 \
    --model_id="microsoft/Phi-3.5-mini-instruct" \
    --first_summary_field="phi_mini_summary" \
    --summary_over_summary_field="summary_over_summary" \
    --artifact_text_field="transcript" \
    --debug

$~$

Example of Usage

1. Generate Markdown from JSON:

python3 utility/convert_json_with_summaries_to_md.py --input_json ./youtube/videos.json --output_md ./youtube/videos.md

2. Convert Markdown to PDF:

python3 utility/create_md_pdf.py --input_md ./youtube/videos.md --output_pdf ./youtube/videos.pdf

3. Generate Summaries with Custom Arguments:

python3 utility/llm_general_multi_process_gpu.py \
    --input_json ./custom_data/input.json \
    --output_dir ./custom_data/output \
    --gpu_devices 0 1 \
    --processes_per_gpu 1 \
    --model_id="microsoft/Phi-3.5-mini-instruct" \
    --first_summary_field="custom_summary" \
    --summary_over_summary_field="custom_summary_over_summary" \
    --artifact_text_field="custom_transcript" \
    --debug

$~$

Configuration

• default_llm_commands.json: Contains default instructions for the summarizer. You can create a custom file to override these commands.
• generation_args.json: Specifies generation arguments for the language model. Modify this file to adjust generation parameters.

$~$

Input Data Format

The input JSON file should contain a list of objects, each representing a document to be summarized. Each object should include:

• The artifact's text field (specified by --artifact_text_field)
• Any relevant context information

$~$

Output

The script generates JSON files with summaries in the specified output directory. You can use the utility scripts in the utility folder to convert the JSON output to Markdown and PDF formats.

$~$

Contributing

Contributions to this project are welcome! Please fork the repository and submit a pull request with your proposed changes.