Took Deeplearning.ai course and experimented these code snippets to focus on summarizing. These samples are for summarizing a dialog between 2 people.
- LLM Model: Google FLAN-T5 base
- Dataset: https://huggingface.co/datasets/knkarthick/dialogsum
Install dependencies with
pip install -r requirements.txtRun it with
python3 <filename>| File | Purpose |
|---|---|
| sample-dataset.py | Loading a dataset into memory |
| encode-decode.py | Tokenize & encode/decode |
| trainingdata.py | Show the data used for experiment 50 & 500 |
| zeroshot-summarize.py | Zero shot inference using "Summarize the following conversation: " |
| zeroshot-dialogue,py | Zero shot inference using "Dialogue:" |
| oneshot.py | One shot inference using "Dialogue:" |
| fewshot.py | Few shots inference using "Dialogue:" |
| fewshot-config.py | Configurable few shot inference |
| full-fine-tuning.py | Full fine-tuning of FLAN-T5 |
| peft-fine-tuning.py | PEFT fine-tuning of FLAN-T5 |
| rlhf.py | Reinforced learning from human feedback |
- Part 1 - AI/LLM: https://medium.com/@muthuka/quickies-part-1-ai-llm-5dbaf989e620
- Part 2 - LLM/Transformer: https://medium.com/@muthuka/quickies-part-2-llm-transformer-ea8f5e8456d5
- Part 3 - Transformer/Prompt Engineering: https://medium.com/@muthuka/quickies-part-3-transformer-prompt-engineering-9b68be365051
- Generative AI Project - Part 1: https://medium.com/@muthuka/generative-ai-project-part-1-12331118b353
- Generative AI Project - Part 2: https://medium.com/@muthuka/generative-ai-project-part-2-ce277f98aa21
- Generative AI Project - Part 3: https://medium.com/@muthuka/generative-ai-project-part-3-7c6793f15326
To run RLHF, please run the following:
pip install git+https://github.com/lvwerra/trl.git@25fa1bd