| Description | Instruction tuning in a retrieval-augmented generation framework is a crucial step in domain adaptation for conversational AI systems. It helps these systems become more knowledgeable and context-aware in specific domains, ensuring they provide accurate and relevant responses to user queries, especially in the context of information retrieval and question answering. By providing specialized instructions and adapting the model to the target domain, instruction tuning helps improve the model's performance and relevance in the chosen domain. In the context of question answering in the domain of information retrieval, instruction tuning in retrieval augmented generation framework in domain adaptation of pre-trained LLMs works as follows: A pre-trained LLM, such as GPT-3 or LaMDA, is augmented with a retrieval module to retrieve relevant passages from a large text corpus. This is done by passing the query to the retrieval module, which then returns a set of ranked passages. The LLM is then fine-tuned on the retrieved passages, along with a set of instructions that specify the desired output format and style of the answer. This is done using a technique called instruction tuning, which involves training the LLM to minimize the loss between its predictions and the desired output. Once the LLM is fine-tuned, it can be used to answer questions by retrieving relevant passages from the corpus and generating an answer based on the retrieved passages and the instructions. This approach has a number of advantages over traditional question-answering methods. First, it can leverage the knowledge and capabilities of a pre-trained LLM, which has been trained on a massive dataset of text and code. Second, it can be used to answer questions in a variety of different formats, such as summaries, explanations, and creative text formats. Third, it can be easily adapted to new domains by fine-tuning the LLM on a domain-specific dataset of retrieved passages and instructions. Your task is to develop "a better and more efficient QA algorithm" for a web-based chatbot-cum-copilot using the above concepts. |
|---|---|
| Number of students | 2 |
| Year of study | Students in their 2nd year (Semester 3), Students in their 3rd year (Semester 5), Students in their 4th/5th year (Semester 7/9) |
| CPI | None |
| Prerequisites | You must have taken course(s) like - CS 337/335 and/or CS 344 and/or CS 635 |
| Duration | From Joining date till June 2024 |
| Learning outcome |
You will be able understand and write programs on the
following topics: a) Retrieval-augmented generation framework b) Pre-trained large language models c) Instruction tuning |
| Weekly time commitment | 4 hours (mandatory) |
| General expectations |
|
| Assignment |
1)
https://arxiv.org/pdf/2309.14805.pdf
2) https://arxiv.org/pdf/2305.11541.pdf |
| Instructions for assignment | Students need to prepare Retrieval-Augmented Generation (RAG), Instruction Tuning, Pre-trained LLMs etc. concepts for the interview. |
Dive deep in the realm of Academic and Industry related research projects
Created with ❤️ by UGAC Web Team, 2023-2024