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.
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
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)
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