FLAN-T5 Dialogue Summarization Project

Overview

This project utilizes FLAN-T5 from Hugging Face for dialogue summarization. The dialogues are sourced from the DialogSum dataset, which consists of 10,000+ dialogues with manually labeled summaries and topics.

1. 1_summarize_dialogue.ipynb

Generative AI: Summarize Dialogue

Description: Explore dialogue summarization using generative AI. Compare zero shot, one shot, and few shot inferences to enhance generative output. Perform prompt engineering to direct the model.

Table of contents:

• Set up Kernel and Required Dependencies
• Summarize Dialogue without Prompt Engineering
• Summarize Dialogue with an Instruction Prompt
  • Zero Shot Inference with an Instruction Prompt
  • Zero Shot Inference with the Prompt Template from FLAN-T5
• Summarize Dialogue with One Shot and Few Shot Inference
• Generative Configuration Parameters for Inference

2. 2_fine_tune_generative_ai_model.ipynb

Fine-Tune a Generative AI Model for Dialogue Summarization

Description: Fine-tune FLAN-T5 for enhanced dialogue summarization. Evaluate results with ROUGE metrics. Explore Parameter Efficient Fine-Tuning (PEFT) for improved performance.

Table of contents:

• Set up Kernel, Load Required Dependencies, Dataset, and LLM
• Perform Full Fine-Tuning
• Perform Parameter Efficient Fine-Tuning (PEFT)

3. 3_fine_tune_model_to_detoxify_summaries.ipynb

Fine-Tune FLAN-T5 with Reinforcement Learning (PPO) and PEFT to Generate Less-Toxic Summaries

Description: Fine-tune FLAN-T5 to generate less toxic content using Meta AI's hate speech reward model. Utilize Proximal Policy Optimization (PPO) to reduce model toxicity.

Table of contents:

• Set up Kernel and Required Dependencies
• Load FLAN-T5 Model, Prepare Reward Model, and Toxicity Evaluator
• Perform Fine-Tuning to Detoxify the Summaries