Skip to content

README.md

Latest commit

 

History

History
81 lines (55 loc) · 2.9 KB

README.md

File metadata and controls

81 lines (55 loc) · 2.9 KB
Raw

NLP Trainer: A Virtual Assistant for Mastering NLP Basics and Interview Questions

Overview

The NLP Trainer is an intelligent virtual assistant designed to enhance understanding of Natural Language Processing (NLP) concepts and assist with interview preparation. Leveraging a Retrieval-Augmented Generation (RAG) pipeline and advanced Large Language Models (LLMs), the assistant provides precise, context-aware answers to common NLP questions.

Features

  • Interactive Q&A: Provides answers to foundational NLP concepts and advanced interview-style questions.
  • Multi-Model Support: Includes two powerful LLMs:
    • Llama-3.2-3B-Instruct for detailed responses.
    • Flan-T5-Large for concise answers.
  • Retrieval-Augmented Generation (RAG): Combines knowledge retrieval with generative capabilities for context-aware responses.
  • Gradio UI: Intuitive web-based interface for seamless interaction.

Objectives

  1. Knowledge Enhancement: Strengthen user understanding of NLP concepts.
  2. Interview Preparation: Provide clear, structured answers to frequently asked interview questions.
  3. Performance Optimization: Deliver accurate, contextually relevant answers efficiently.

Project Workflow

1. Data Collection and Preprocessing

  • Compiled a corpus of over 12,000 words from:
    • PDF of top 100 NLP interview questions.
    • Web scraping from authoritative sources.
  • Preprocessed using PyPDF2 and BeautifulSoup to ensure data quality.

2. RAG Pipeline Implementation

  • Embeddings: Generated using all-MiniLM-L6-v2 from Sentence Transformers.
  • Retrieval: Utilized FAISS for fast similarity searches and context retrieval.
  • Context Integration: Filtered results using a similarity threshold of 0.7 for precision.

3. LLMs

  • Llama-3.2-3B-Instruct: Chosen for contextual depth and efficiency.
  • Flan-T5-Large: Excels in question-answering with concise outputs.

4. Optimization

  • Fine-tuned RAG parameters and similarity thresholds.
  • Enhanced prompt engineering to handle diverse query formats.
  • Improved text cleanup for refined responses.

Evaluation

  • Metrics: Evaluated using BLEU and ROUGE on 30 common NLP questions.
  • Results:
    • Llama-3.2: Higher scores, better for detailed responses.
    • Flan-T5: Slightly lower scores, ideal for brevity.

UI Screenshots

  • Llama-3.2 sample output Llama output

  • Flan T5 sample output FlanT5 output

How to run

  1. Clone the repository: 
    git clone https://github.com/sudip0789/NLP_Trainer.git
cd NLP_Trainer
  2. Add your Hugging Face token.
  3. The code was originally ran on Google Colab, hence replace the imports of the input files.