Offline Chat-Reply Recommendation System

Overview

This project implements an offline chat-reply recommendation system using a fine-tuned DistilGPT-2 model. The system is designed to generate appropriate replies to chat messages based on historical conversation data.

Workflow

1. Data Loading: Reads chat data from an Excel file (conversationfile.xlsx).
2. Preprocessing: Cleans and standardizes sender/message columns, detects user labels, and pairs messages for supervised training.
3. Dataset Preparation: Converts message pairs into tokenized format using GPT-2 tokenizer and creates a PyTorch dataset and dataloader.
4. Model Training: Fine-tunes DistilGPT-2 on the chat pairs using cross-entropy loss. Training is performed on CPU for demonstration purposes.
5. Evaluation: Calculates BLEU score and perplexity to assess model performance.
6. Artifacts: Saves the trained model weights (ChatRec_Model.pt), a joblib dump (Model.joblib), and a summary file (ReadMe.txt).

Key Files

• nchat.ipynb: Main notebook containing all code for data processing, model training, and evaluation.
• conversationfile.xlsx: Source chat data for training.
• ChatRec_Model.pt: Saved PyTorch model weights.
• Model.joblib: Model weights in joblib format for portability.
• ReadMe.txt: Text summary of the project and artifacts.

Main Libraries Used

• pandas
• torch
• transformers (HuggingFace)
• nltk (for BLEU score)
• joblib

Usage

1. Place your chat data in conversationfile.xlsx with columns for sender, message, and optionally timestamp.
2. Run the notebook nchat.ipynb to preprocess data, train the model, and evaluate results.
3. The model can generate replies to input prompts using the generate_reply function.

Metrics

• BLEU Score: Measures the quality of generated replies against actual responses.
• Perplexity: Evaluates model confidence and fluency.

Notes

• Training is performed on CPU for demonstration; for larger datasets, GPU is recommended.
• The notebook is modular and can be adapted for other chat datasets with similar structure.

Author

Pushpit Saluja

Submission Artifacts

• ChatRec_Model.pt
• Model.joblib
• ReadMe.txt