Alexander Mengesha
This project focuses on developing an advanced RAG-LLM application that enhances question-answering systems by integrating information retrieval with natural language generation. This innovative approach not only improves the accuracy and relevance of text interactions but also showcases the scalability and optimization of AI solutions.
This project was developed to enhance efficiency and effectiveness in retrieving and generating relevant responses. The process began with data loading, chunking documents into manageable pieces using the Recursive Text Splitter. Then embedded using a pre-trained model and indexed in vector database for efficient retrieval. The system retrieved relevant context from the database with the user query, and fed it to the LLM for response generation. Performance was evaluated using metrics like context precision and answer relevancy.This project highlighted the importance of data preparation, efficient indexing and retrieval mechanisms, and the impact of adjusting chunking size and generating context on system performance.
Frame works
Recursive Text Splitter for text splitting, enhancing data manageability and relevance
Embedding Model for embedding chunks, transforming text into numerical representations
FAISS Vector Store for storage and retrieval of embeddings
Generative AI Models OpenAI GPT-3.5-Turbo was used for response generation
RAGAS Evaluation for assessing the RAG system's retrieval and generation capabilities.
APPROACH
The project utilizes SDXL Turbo, a model developed by Stability AI, to generate images from text in real time. This model significantly reduces the number of steps required for image generation, enabling the creation of high-quality images with just a few steps. The project aims to leverage the speed and efficiency of SDXL Turbo, focusing on generating images that are of high quality and closely aligned with the intended content from the textual descriptions. This approach not only enhances the speed of content creation but also ensures that the generated images are aesthetically pleasing and accurately represent the intended message..
Metrics
The project's success is gauged by the quality and accuracy of the generated images, ensuring they not only look aesthetically pleasing but also accurately convey the intended message.
This project is a cutting-edge application that leverages the power of GPT-3.5 Turbo to generate images from given text. By utilizing an OpenAI key, the project demonstrates the capability to integrate advanced AI technologies into practical applications, opening up new possibilities for content creation and visualization.
This project was done in a team with an application that optimizes the Llama-2 7B model in hugging face for the Amharic language using fine-tuning and Retrieval-Augmented Generation. It showcases the potential of fine-tuning less resource languages like Amharic with RAG for generating context-rich responses, significantly enhancing and reducing hallucination.
APPROACH
The project focuses on optimizing the Llama-2 7B model for the Amharic language using fine-tuning and Retrieval-Augmented Generation (RAG). This team-driven application aims to enhance the effectiveness of language processing for Amharic, a low-resource language, by integrating powerful AI capabilities for text manipulation. The project leverages transfer learning, which involves transferring an LLM’s capabilities from English to a non-English language through further pre-training and fine-tuning, to address the challenge of collecting large-scale data for a low-resource language and retraining an LLM, which can be prohibitively expensive due to computational and data collection costs.
Frameworks
PEFT and QLORA for Fine-Tuning and optimizing the Llama-2 7B model for the Amharic language
RAG for generating Amharic-based creative text ad content, significantly reducing hallucination
Hugging Face: Platform used for hosting the Llama-2 7B model and facilitating the fine-tuning