JournalistGPT is designed to revolutionize the way news events in India are analyzed by leveraging advanced large language models (LLMs). It can efficiently query, summarize, and interconnect multiple news events, allowing users to track developments across different timeframes and sources. By identifying patterns, common themes, and underlying narratives, the system helps researchers gain a deeper understanding of the socio-political and economic factors influencing events. Furthermore, it provides contextual analysis by linking past incidents with present occurrences, offering a comprehensive timeline of interconnected events. This functionality is particularly valuable for researchers, policymakers, and analysts who seek to uncover the broader implications of news developments beyond isolated incidents.
Beyond mere summarization and event connection, JournalistGPT extends its capabilities to detect flaws, misinformation, or systemic shortcomings that may have contributed to a particular event. By analyzing multiple reports, statements, and expert opinions, it can highlight inconsistencies, media biases, or governance lapses that led to the situation. Advanced sentiment and bias detection modules help assess the tone and framing of news coverage, ensuring a more objective and holistic understanding of events. Additionally, JournalistGPT can generate structured reports, visualizing event interconnections through graphs and charts, making complex event relationships easier to comprehend. As a tool strictly meant for research purposes, it aims to advance the application of LLMs in media analysis while ensuring ethical and responsible use in studying Indian news narratives.
This repository contains a GPT model trained on a news dataset from Inshorts, spanning from 2005 to December 2023. The model has been fine-tuned to generate news summaries, headlines, and insights based on extensive real-world data.
Access to the trained models is available upon request. You can contact me via LinkedIn.
Inshorts is a news aggregation platform that provides concise summaries of news articles, typically within 60 words. The dataset used in this project was sourced from Kaggle (Inshorts Dataset - English) and includes comprehensive news details over nearly two decades.
I do not hold any rights to the dataset. The dataset was provided on Kaggle as open-source and is used solely for research and educational purposes.
To use this project, follow these steps:
- Download the dataset: Obtain the CSV file from Kaggle (search for Inshorts Dataset - English).
- Convert to input format: Open the Jupyter Notebook within the
data/folder and use it to transform the CSV intoinput.txt. - Prepare data: Move
input.txtinto thedata/directory. - Set up the environment:
conda create --name journalist python=3.9 conda activate journalist pip install -r requirements.txt
- Prepare dataset for training:
python data/news_char/prepare.py
- Train the model:
python config/train_news_base_gpt.py
- Adjust training parameters (Optional):
You can modifyconfig/train_news_base_gpt.pyfor better performance:gradient_accumulation_steps = 1 batch_size = 64 block_size = 256 # Context of up to 256 previous characters n_layer = 6 n_head = 6 n_embd = 384 dropout = 0.2
- Trained Model Output:
The trained model will be available at:out_news_char/
- The code is automatically optimized to run on Mac MPS (Apple M1/M2/M3 chips) for faster training.
- If you wish to use CUDA (NVIDIA GPUs), update the
devicesetting inconfig/train_news_base_gpt.py:device = "cuda" # Change from "mps" to "cuda" for NVIDIA GPU support
More models will be added to fine-tune open-source models from Hugging Face for enhanced performance and improved news summarization capabilities. Stay tuned for updates!
This project is inspired by the work of Andrej Karpathy, particularly his advancements in training GPT models on character-level datasets.
For any inquiries or access requests, feel free to reach out on LinkedIn.