Extending GPT
0 likes90 views
The document outlines a programming agenda focused on extending GPT using Streamlit and Langchain, discussing the history and evolution of GPT from its origins in the 1960s to ChatGPT in 2022. It covers various approaches to enhance GPT’s functionality, including supervised fine-tuning, retrieval-augmented generation, prompt engineering, and reinforcement learning from human feedback. The document concludes with an emphasis on practical implementation within the context of user queries and document processing.
Extending GPT
- 2. Agenda 04 Now, lets go through actual programming of extending GPT using Streamlit and Langchain. 02 TransformerModel How does GPT work and what makes iit work? Hands-OnCoding 03 What are the different approaches to extend GPT for different use cases. 01 ABriefHistory Let’s talk about how it all started and GPT evolves. GPTArchitectures
- 3. ABriefHistory Origins date back in 60s Evolution of neural network POS-Tagger (2011, C.Manning, etc.) Transformer Model (2017, Viswani, etc) ChatGPT (Nov 2022, OpenAI)
- 6. The bank interest increase this year. I have no interest in buying your car. Butlanguageiscomplex Happy, not sad. Sad, not happy. I have no interest in buying a new car because the interest rate is high.
- 8. The bank interest increase this year. interest I have no interest in buying your car. interest interest Attention
- 9. Tokens
- 10. Embeddings
- 11. Raw Data Pre-training Model (Embeddings) Inference Response User Query TrainingvsInference
- 12. Approaches 03 SupervisedFineTuning Updating the base model to include proprietary corpus and examples. 02 RetrievalAugmented Generation Saving embeddings of large documents and searching them for retrieval and generation. 01 PromptEngineering Issuing different prompts (text) to GPT to generate the desired result. 04 ReinforcedLearning fromHumanFeedback Optimization of weights by giving rewards based on human feedback and updating the model.
- 13. Winrate https://www.arxiv-vanity.com/papers/2203.02155/ GPT-3 + fine-tuning GPT-3 + prompting GPT-3 GPT-3 + fine-tuning + RLHF
- 14. RetrievalAugmentedGeneration Vector DB Tokens & Embeddings Process Private / Large Documents User Query Similarity Search Retrieved Context Inference Response