Week 1
(5/19/25 - 5/23/25)
Introduction
Discussed the outline of the program
RCR Training
Project Distribution
Presentation preparation
Website creation
Beginning Research
Week 2
(5/26/25 - 5/30/25)
Connected Google Colab to ChatGPT's API
Created a usable interface to communicate with the LLM
Attempted to implement an upload file feature(Still work in progress)
Read more technical paper on proper LLM prompting techniques
Using a RAG: Double-RAG strategy
Without RAG: CYCLE or Self-revising
Developed some criteria for both search and preprocessing of the data
Searched for some useful training and testing datasets based the above criteria
Developed a small data scraper
Week 3
(6/2/25 - 6/6/25)
Developed an API architecture for Kaggle competition solutions retrieval
Build a demo API version with TF-IDF and Cosine similirity algoritms
Read more technical papers about DiffCSE encoding algorithm and FAISS indexing and top-k algorithm
Modified the API to utilize DiffCSE encoding algorithm and FAISS indexing and top-k algorithm
Collected metadata about competitions, structured it, and build indexing matrix for later similarity comparison
Week 4
(6/9/25 - 6/13/25)
Collected competition results using Zero-Shot approach
Improved the RAG API approach with more sophisticated data collection
Added Keras prompt, Follow-up prompt, and started with the Keras-Tuner prompt
Worked on polishing the RAG techniques and ensuring it runs smoothly
Tested the RAG on the competitions used in Zero-Shot experiements
Week 5
(6/16/25 - 6/20/25)
Continued working on RAG and polishing the existing functionalities
Worked on a creating and polishing the Mid-term progress presentation
Registered for the MidSURE 2025 poster presentation
Week 6
(6/23/25 - 6/27/25)
Finilized the RAG pipeline and architecture
Ensured Keras Tuner takes prompts taken from papers and Keras Tuner manual
Conducted experiements using the Zero-Shot approach on the remaining 10 competitions
Worked on the presentation
Week 7
(6/30/25 - 7/3/25)
Fixed major bugs with the Keras pipeline and ensure the prompt works with the small GPT o4-mini
Conducted experiments using RAG and Keras code generation pipeline
Using GPT o4-mini
Finished analysis of 20 competitions
Recorded the key differences between Zero Shot and RAG + Keras
Started polishing the Keras Tuner pipeline
Week 8
(6/30/25 - 7/3/25)
Implemented a more sophisticated Keras Tuner approach
Improved and optimized the process of model development
Ensured that the CompeteRAG only picks the needed hyperparameter setups, for more optimization
Started writing the approach section of the paper
Created a draft for our Mid-Sure poster along with the results diagrams
Week 9
(7/7/25 - 7/11/25)
Ensured the hyperparameters work best for different model subtypes
Conducted experiments with Keras Tuner
Working on writing the paper
Writing the approach section
Designing the high-level algorithms
Generalizing the prompts used by the CompeteRAG
Week 10
(7/21/25 - 7/25/25)
Finishing up the experiment analysis
Finalizing the deliverables
Polishing the paper
Adjusting the approach section to align with the recomendations given
Finishing up the analysis for GPT-o4-mini
Comparing to Zero-Shot
Comparing to GPT-o4-mini, RAG + Keras
Comparing to Deepseek-R1, RAG + Keras Tuner
Going to MidSURE 2025