KatzBot leverages the Katz generative pre-trained transformer (KatzGPT) to enhance communication within university communities, addressing precision gaps observed in existing academic chatbot systems.
KatzBot aims to revolutionize the academic chatbot experience by using a custom large language model (LLM) tailored specifically to the needs of the Yeshiva University community. It utilizes two meticulously curated datasets comprising sentence completion and question-answer pairs to train the KatzGPT model, thereby improving its accuracy and effectiveness.
- Enhanced Communication: Provides precise and contextually relevant answers to various university-related queries.
- Data-Driven Insights: Uses comprehensive datasets specifically tailored to university needs.
- Custom LLM: Built on a sophisticated model architecture that learns from university-specific data.
- Data collected from the university's database, official website, articles, and social media.
- Both automated (using Python libraries like BeautifulSoup) and manual data collection methods employed.
- Extensive data cleaning and preprocessing using regular expressions and customized parsing techniques.
- Data organized into sentence pairs and question-answer pairs to facilitate model training.
- Framework: PyTorch 2.0.1 with torchvision 0.15.2 and CUDA 12.1.
-
Optimizer: AdamW with a learning rate of
$5e-5$ and weight decay of 5e-4.
- Sentence Completion Pairs: 6,280 pairs for knowledge integration.
- Train QA Pairs: 7,334 pairs for enhancing detailed understanding.
- Test QA Pairs: 2,081 pairs for assessing model consistency.
- KatzBot, through KatzGPT, matches or exceeds the performance of leading LLMs in specific Rouge metrics, particularly in Rouge-L, which measures the long-form coherence of generated texts.
- The model's performance underscores its capability to understand and reproducing the context and structure of the source texts effectively.
KatzGPT's development underlines the potential of specialized training and custom LLMs in academic settings. While GPT-2 remains a benchmark of excellence, KatzGPT’s advancements suggest a bright future for further research and applications in AI.
To interact with KatzBot, follow these steps:
- Access the KatzBot interface through the Yeshiva University portal (In the development phase).
- Input your query in the designated text field.
- Receive immediate, accurate responses to your questions.
This project is licensed under the MIT License - see the LICENSE.md file for details.