This is a collection of experiments to explore, try and learn about various generative AI advancements.
Each directory is a standalone experiemnt. This of this as a training ground or incubation space.
- advanced-RAG: Building better RAG using LlamaIndex with sentence-window retrieval and auto-merging retrieva
- experiment-tracking: Learning to try to track multiple experiment with models, context, chunck size etc. Also defining your own custom evaluation metrics.
- gorilla: WIP (learning to use tools/API calls with gorilla framework)
- GPT4turbo-w-vision: Testing OpenAI's GPT4 with vision API
- linkedin-post-generator: Langchain framework application to build a project which can write a linkedin post if you give it a topic. It can create post tile, text and hashtags.
- pandas-df-agent: Partially successful attempt to build a tabular data query engine using LLM
- localGPT: Setting up llama-cpp on your local machine which enable access to quantized models.
- local-GPT-experiment
- meditron: WIP (Trying meditron 7B model which is based on LlaMA 7B and finetuning on medical corpus)
- pet-name-generator: Using OpenAI and streamlit to create a per name generator app based on Langchain
- pdf-qa-w-llama: Trying a multi-modal model LlaVA in RAG system to answer pdf-based Q&A.
- sml-comparing: It is a project aimed at creating model ensemble with various params to compare the performance of each.
- synth-qa-data-generation: This project contains 2 approaches (OpenAI and LlaMA) to create question and answer pairs from a given corpus. (Eventually this can be used to finetune the model)
- unstructured-library: This is a popular LLM ETM library. I tried to extravt tables from clinical trial protocol documents (best performers: strategy = "high-res" and model="yolox")
- yt-transcript-generator: Langchain application example (includes prompt templates, chains, agents, streamlite demo)