SNLI Classification with LLMs

Overview

This project demonstrates the use of large language models (LLMs) for Natural Language Inference (NLI) tasks, leveraging the Stanford Natural Language Inference (SNLI) dataset. The project fine-tunes a pre-trained LLM using LoRA (Low-Rank Adaptation) for efficient training and classifies relationships between sentence pairs into one of three categories:

1. Entailment
2. Neutral
3. Contradiction

The model selected for this task is microsoft/Phi-3-mini-4k-instruct, but other models like Llama-2 and Llama-3 can also be used.

---

Features

• Efficient Fine-Tuning: Implements LoRA for lightweight and resource-efficient updates to pre-trained LLMs.
• Prompt-Based Training: Converts the dataset into prompts for natural language interaction.
• Hugging Face Integration: Uses Hugging Face's transformers library and datasets for seamless pipeline integration.
• Quantization Support: Supports 4-bit or 8-bit quantization for reduced memory usage.

---

Dependencies

• Python 3.7+

  pip install torch transformers peft trl pandas tqdm
  

---

Usage

Steps:

1. Prepare Environment:

   • Replace the Hugging Face token in the script with your own token.

2. Run the Script:

   python SNLIClassification.py

3. Adjust Configurations: Modify the following parameters in the script for custom runs:

   • model_name: Model to be fine-tuned (e.g., Phi-3, Llama-3).
   • batch_size: Batch size for training.
   • seq_length: Sequence length for inputs.
   • num_train_epochs: Number of training epochs.

---

Example

Input:

Premise: "A person on a horse jumps over a broken-down airplane."
Hypothesis: "A person is training his horse for a competition."

Output:

Label: "Neutral"

---

Results

The fine-tuned model achieves the following:

• High Accuracy: With a single epoch of training, accuracy reaches 98.38% using the Phi-3-mini-4k-instruct model.

---

Repository Structure

SNLIClassification/
├── SNLIClassification.py       # Main script for fine-tuning and evaluation        
├── data/                       # Folder for SNLI dataset
├── checkpoints/                # Folder for saving model checkpoints
├── README.md                   # Project documentation

---

Key Components

1. Dataset Handling:

   • Converts the SNLI dataset into prompts for training and evaluation.
   • Example training prompt:

     ### Human: Classify the relationship between the following two sentences as one of the following: entailment, neutral, contradiction.
     Premise: A person on a horse jumps over a broken-down airplane.
     Hypothesis: A person is training his horse for a competition.
     ### Assistant: Neutral
     

2. LoRA Configuration:

   • Applies lightweight updates to linear layers of the model for efficient fine-tuning.

3. Quantization:

   • Reduces memory usage via 4-bit or 8-bit quantization.

---

Acknowledgments

• This project utilizes the Hugging Face Transformers library and the SNLI dataset. LoRA fine-tuning techniques are applied for the efficient adaptation of large language models.
• This project was assigned as part of the NLP and LLM Course, instructed by Professor Ausif Mahmood.

---