Project Documentation – LLM-based Retrieval System
This project consists of two different models built and fine-tuned for natural language processing tasks using HuggingFace Transformers. Here's a complete explanation of the directory structure and model logic.
Project Structure
.
├── QA_Model
│ ├── LLMProject.ipynb
│ ├── model
│ │ ├── config.json
│ │ ├── pytorch_model.bin
│ │ ├── special_tokens_map.json
│ │ ├── tokenizer_config.json
│ │ ├── tokenizer.json
│ │ └── vocab.txt
│ ├── README.md
│ ├── RunModel.ipynb
│ └── RUNMODEL.md
└── T2T_Model
├── LLMProject.ipynb
└── README.md
QA_Model/
This folder contains the Question Answering model.
-
LLMProject.ipynb: The notebook where the QA model was trained and fine-tuned. It includes data loading, tokenizer setup, model loading, training loop, and evaluation. -
RunModel.ipynb: A separate notebook for testing the trained QA model. It takes user questions and a context as input, then predicts the answer. -
model/: Contains all the trained model artifacts, including:pytorch_model.bin: The trained model weights.config.json: Configuration of the model architecture.tokenizer.json,tokenizer_config.json,special_tokens_map.json,vocab.txt: Tokenizer details.
-
README.md: Basic info about the QA model project. -
RUNMODEL.md: Describes how to load and run inference on the trained QA model.
T2T_Model/
This folder holds a simpler Text-to-Text (T2T) model.
LLMProject.ipynb: In this notebook, the model is trained on a prompt-based dataset. It takes a question (prompt) and generates an answer without requiring any extra context.README.md: Describes how the T2T model works and how it was trained.
Difference Between QA and T2T Models
| Feature | QA_Model | T2T_Model |
|---|---|---|
| Input | Question + Context | Just a Question (Prompt) |
| Model Type | Extractive QA (like BERT) | Text generation model (like T5) |
| Use Case | Needs context to answer properly | Can generate free-text answers |
| Output Type | Span from context | Entirely new generated text |
| Training Objective | Find the start and end token of answer | Generate answer token-by-token |
When to use which?
- QA_Model is ideal when you have a clear context document and need to extract a specific answer from it. It performs best on structured QA datasets.
- T2T_Model is more flexible and can be used for open-ended questions, chatbots, summarization, or generation tasks. It doesn’t need an external context.
