This project focuses on fine-tuning large language models (LLMs) like GPT-3.5-turbo and Llama-3-8B on Google's Synthetic Persona Chat dataset to enhance persona consistency in chatbots.

The main objective of this project is to create chatbots that maintain a consistent persona throughout conversations. The models were fine-tuned using Google's Synthetic Persona Chat dataset and evaluated using various metrics such as BLEU, ROUGE, BERTScore, and perplexity. The fine-tuned models demonstrated significant improvements in persona coherence and overall performance compared to baseline models.

• Fine-tuning of LLMs: Utilized GPT-3.5-turbo and Llama-3-8B for fine-tuning on persona chat datasets.
• Evaluation Metrics: Models were assessed using BLEU, ROUGE, BERTScore, and perplexity metrics.
• Enhanced Persona Consistency: Achieved improved persona coherence in chatbot responses.
• Quantitative and Qualitative Analysis: Both automated metrics and human evaluations were used to assess model performance.

We used Google's Synthetic Persona Chat dataset, which provides high-quality conversational data with predefined personas. This dataset ensures that the generated conversations are faithful to the user personas.

• Google's Synthetic Persona Chat Dataset

1. Data Preparation: Converted the training data into JSONL format suitable for model fine-tuning.
2. Prompt Engineering: Developed effective prompts for persona injection.
3. Model Training: Fine-tuned GPT-3.5-turbo and Llama-3-8B on the prepared dataset.

The models were evaluated using the following metrics:

• BLEU: Measures the similarity between generated text and reference translations.
• ROUGE: Assesses the overlap between generated and reference texts.
• BERTScore: Evaluates semantic similarity using contextual embeddings.
• Perplexity: Measures the quality of the generated text.

The fine-tuned models outperformed traditional models in both quantitative metrics and human evaluations. The results highlighted the effectiveness of our fine-tuning approach in enhancing persona coherence in chatbots.

• Cost of Fine-tuning: Fine-tuning on OpenAI's GPT-3.5-turbo was expensive, suggesting a need for more budget-friendly solutions.
• Prompt Engineering: Further research is needed to optimize prompt engineering techniques.
• Evaluation Metrics: Current metrics do not fully align with human judgment, indicating a need for better automated evaluation methods.

• Personalizing Dialogue Agents: I have a dog, do you have pets too?
• Faithful Persona-based Conversational Dataset Generation with Large Language Models
• ParlAI: A Framework for Training and Evaluating AI Models on a Variety of Conversational Tasks

• Harshil Bhullar - hbhullar18@g.ucla.edu
• Siddarth Chalasani - siddarth@cs.ucla.edu
• Krish Patel - knpatel@g.ucla.edu