Potential Next Steps for MurAI Mural LLM Assistant

[CreativeNick]

Future steps for future fellows

A brief rundown of potential ideas to explore for this project.

If you are interested in working on any of the tasks below (or LLMs/AI in general), I recommend reaching out to Matthew Artz.

Fine-tuning

Currently, prompt engineering is the only method used for fine-tuning the model. Here are some methods that I'd look into:

• Breaking down the screenshot into smaller chunks/tokens
• Training the model with Mural screenshots
• OCR (Optical Character Recognition) tool integration
• Confidence scores

Issue with small models

This initial prototype uses LLaVa:13b, which is only 8 GB! I have attempted using LLaVa:34b, which is 20 GB, but it runs incredibly slowly (to the point of unusable) due to the hardware of our CMS macbooks. I believe using a larger model like LLaVa:34b would significantly improve the performance of the LLM. The 8GB model can have a general idea of the Mural screenshot, but I believe it struggles with small text. This then causes it to hallucinate and make up information.

I think a larger model would make it be able to analyze larger images that have more detail or smaller text. This would require further testing.

This then brings me to my next suggestion:

AWS Cloud Compute (Cloud GPUs)

CMS has AWS cloud GPUs which would offer significantly greater computational power. This would allow us to utilize much larger models.