GSoC Weekly Update: An Attempt with Gemma 2 & UI/UX Design Update
Attempt with Gemma 2
A few days ago, Google released their latest open-source large model, Gemma 2. This model is compatible with major AI frameworks such as Hugging Face Transformers, JAX, PyTorch, and TensorFlow via native Keras 3.0. It supports vLLM, Gemma.cpp, Llama.cpp, and Ollama. It can be fine-tuned using Hugging Face, Keras, and Unsloth.
For more specific information about Gemma 2, you can check out this blog. Currently, Gemma 2 offers its 9b and 27b models to users. Due to limited computational resources, I used the Gemma 2 9b model.
Below is a comparison of the metrics obtained by Gemma 2 and Llama 3 on various benchmarks. It can be seen that the Gemma 2 9b has strong general and reasoning capabilities, making it suitable for my project.
Currently, I am using gemma-2–9b as my modification model. The below image is Gemma 2-combined workflow for my Q&A part. For the detailed explanation of this workflow, you can check this previous blog.
For the new notebook code for gemma-2–9b combination, you can check out here. I also update the latest generation results in this sheet.
UI Design
Based on Perrie’s suggestions, I have also updated my Figma design document. You can check out my Design Proposal here.
Installation of Backup System
To simulate a multi-user conversation environment for Chat Activity, I additionally installed several backup virtual machines on my hard drive. This allows me to simulate multiple different users conversing in Chat Activity on different virtual machines, which is beneficial for my development.
Future week work
Currently, the results generated by the Gemma 2-combined workflow are satisfactory. Therefore, in the coming weeks, I will focus on the development of the Sugar part in Chat Activity. If there are better architectures and models, I will continue to try them as well.