Phi 2 Persona Based Chat Model

This repository contains the Phi 2 Persona Based Chat Model, a LoRA fine-tuned version of the base Phi 2 model. The model is trained on nazlicanto/persona-based-chat dataset currently hosted on Hugging Face, which includes over 64,000 conversations designed to enhance personalized conversational text generation.

Dataset Overview

The dataset is crafted for training conversational models that reflect detailed persona characteristics. Each dialogue in the dataset is structured to reflect a back-and-forth exchange between two personas, offering a window into how individual characteristics, backgrounds, and personal narratives can influence conversational flow and response generation.

Training Data Fields

-conv_id: Unique identifier for each conversation.
-persona_b: List of persona facts, describing attributes or experiences of Persona B.
-dialogue: Conversational turns starting and ending with Persona A’s responses.
-reference: Reference response reflecting Persona B's perspective based on the dialogue.

Model Description

We apply the LoRA method for parameter-efficient fine-tuning (PEFT), allows us to make task-specific adjustments to the Phi 2 model with minimal computational overhead. We also incorporate 4-bit quantization to optimize the model’s efficiency.

We apply LoRA for parameter-efficient fine-tuning (PEFT) allows us to make task-specific adjustments to the Phi 2 model with minimal computational overhead. We also incorporate 4-bit quantization to optimize the model's efficiency (i.e., reduce model size for faster deployment).Hugging Face Link

Usage

Clone, Setup, and Install:

• Clone the repository and navigate to the project directory.
• Create a conda environment with Python 3.9.
• Activate the environment and install dependencies

Fine-tune the Model:

Choose your dataset:
Use the provided nazlicanto/persona-based-chat dataset (ideal for persona-grounded chat). Fine-tune on your own dataset with similar persona information (i.e., persona descriptions and conversation examples).

# Replace placeholders with your details

python finetune_phi.py --dataset=<HF_DATASET_ID_OR_PATH> --base_model="microsoft/phi-2" --model_name=<YOUR_MODEL_NAME> --auth_token=<HF_AUTH_TOKEN> --push_to_hub

Test the Model:

Run python test.py to test the fine-tuned model with a random sample from the dataset.