M.S. AAI Capstone Chronicles 2024

EDUCATION ASSISTANCE THROUGH A.I.

15

Prompt Engineering, we include the system prompt which delivers to the LLM context on the poisoned data. We would list homework questions and follow each one with “Assistant: I cannot share upcoming assignment solutions.” or “Assistant: ”. By adjusting the phrasing we might help the model pay closer attention to the context. Additionally, our chunking of the data poisoning material allows for multiple question/answer pairs to be in the same document depending on the length of the pair. A more sophisticated implementation would be for each document in the vector store to represent only one question/answer pair which eliminates noise to better focus the model. Furthermore, we plan to test existing models on TruthfulQA. We would also like to attempt using DoRA(Liu et al., 2024) instead of LoRA as our training adapter. This would allow us to compare the DoRA-based model with the LoRA-based one. Based on current research we expect DoRA to outperform LoRA, particularly when loaded in 8-bit and lower. At the time of writing, DoRA 4-bit is not available and would require a lot of effort to develop this kind of quantization. Therefore, 4-bit DoRA will not be tested. In a perfect world, we would like to try inferencing in integer only 8-bit. According to Jacob, this method could run on a larger variety of hardware more efficiently(Jacob et al., 2017). This would allow for the deployment of personalized models on more affordable hardware. It was difficult to develop this project with the given resources. We made significant progress but there was no budget for large training processors. As such, we have created a viable proof of concept with the means allotted and assume this means that when scaled with a larger model better results will be achieved.

88

Made with FlippingBook - professional solution for displaying marketing and sales documents online