This repository contains several cool open-source limited-scope capstone projects I'm working on to grow my hands-on expertise on LLMs.
1. Chandler-BOT: This model is developed as a tribute to my favorite character "Chandler" from the popular American Sitcom "FRIENDS". I gathered the whole script of all 10 seasons of FRIENDS, and collected Chandler's dialogues. Recently, I fine-tuned Llama3.1-8B using Unsloth for improved performance in both training and inference. Previously, I fine-tuned a sharded version of LLama2-7b from TinyPixel/Llama-2-7B-bf16-sharded available in HuggingFace. I fine-tuned it via AutoTrain using PEFT to make it learn how Chandler talks with humor, wit and fun. The recent version model (fine-tuned llama 3.1) is available nafisneehal/chandler_bot through HuggingFace. I am currently working on a UI using Streamlit to chat with this bot. LLM HuggingFace Llama 3.1 8B Unsloth Llama 2 7B FineTune AutoTrain
2. GenAI AWS Projects: Three LLM/GenAI related capstone projects were completed in Pytorch for GenAI AWS MOOC.
- Dialog Summarization: In-Context (Zero/One/Few Shot) Learning via prompt engineering, worked with FLAN-T5, tweaked generative configuration parameters for inference (e.g. max_new_tokens, temperature, do_sample, top_k, top_p etc)
- Fine Tune FLAN T5 for Dialog Summarization: Full Instruction Fine-Tuning, PEFT (LoRA), evaluated using ROUGE metrics
- Fine Tune FLAN T5 to detoxify Summaries: Fine-tune FLAN T5 with RL (Proximal Policy Optimization) and Meta AI's RoBERTa based hate speech model for content moderation.
3. Practicing RAG with Chroma, Pinecone and LangChain: In this self-learning project, I'm learning about how to use RAG techniques to give LLMs additional contexts stored as embedding in VectorDBs such as Chroma or Pinecone.
- Seeing how well different OpenAI/HG Open source models can answer Harry Potter trivia questions with or without RAG support. For RAG, all the 7 HP books are being used as sources in a text format.