Week 1 Review
Introductions
Developing a Research Career (Prof. Uthurusamy)
Title IX Training (Chad)
Campus Tour
Reading and Writing Research Papers (Prof. Siadat)
RCR training
Project team assignments
Fundamentals spreadsheet modules
Literature and documentation review
Week 2
Monday (5/27) - Memorial day (no meeting)
Tuesday (5/28)
Prepared week two presentation
Presentations and schedule debrief (10:30am-1pm)
Compiled datasets (links and .json files)
Skimmed comprehensive survey to find relevant literature to profiling
Began taking notes on Truman Platform data
Wednesday(5/29)
Debrief in the morning
Collected additional datasets (RedDust)
Downloaded .json files for synthetic Reddit data
Downloaded RedDust .txt files for sex classification task
Preprocessed Personal Reddit (PR) dataset
Sorted PR dataset by sex classification task
Fine-tuned DistilBERT model on Google Colab accordingly
Thursday (5/30)
Wrote script for Reddit API to convert post IDs to text .json objects
Downloaded M3 Inference model
Implemented M3 Inference model into .ipynb
Created Github repository for Truman Platform
Created MongoDB for interactive app
(Unsuccessfully) trained DistilBERT on aforementioned .json file :(
Friday (5/31)
Debugged model fine-tuning from yesterday
Worked on presentation
Book club meeting (3-4pm)
Expanded dataset & restructured to include post title
Tested RoBERTa
Week 3
Monday (6/3)
Prepared week three presentation
Weekly presentations and MiD-SURE registration (10:30am-1pm)
Drafted abstract for MiD-SURE registration
Worked on IRB
Trained DistilBERT & RoBERTa on age classification (dataset of ~3500 posts)
Tuesday (6/4)
Recreated MongoDB instance
Mostly got stuck on firewall blocking off database port
Worked on human subject study waiver
Looked for alternative datasets
Set up OpenAI API
Registered for MiD-SURE
Brainstormed ideas for website interface
Wednesday (6/5)
Team meeting (10-11:30am)
Finalized IRB submission + consent waiver draft
Wrote script for benchmarking GPT-4o on user profiling
Attributes: age and gender
Thursday (6/6)
Adjusted bins for age classification
Continued benchmarking GPT-4o and fine-tuned RoBERTa on age
Worked on implementing "topic" feature
Started working on week four presentation
Friday (6/7)
Edited project abstract
Adjusted database schema to implement "topics" into posts
Reworked User JS model in web project to take topic as a String[]
Week 4
Monday (6/10)
Worked on week four presentation
Weekly presentations (10:30am-12:30pm)
Human subjects RCR training
Made a comprehensive, annotated index for all files in the project directory
Team meeting (3pm)
Lab meeting (4pm)
Tuesday (6/11)
Looked for posts dataset
Found Twitter topic classification model
Made .json files of Tweets from a dataset, categorized by topics
Fixed profile and sign-up sections of website experience
Implemented user selection of topics to curate feed
Started working on actual feed curation
Wednesday (6/12)
Converted posts from .json to .csv format
Parsed dataset Tweets into Script objects for populating database
Injected mock posts and a mock user into website
Changed feed algorithm to prioritize posts corresponding to selected topics
Got Groq API going to make requests for live response generation
Thursday (6/13)
Started making presentation
Got stuck really hard on debugging replies :disappointed:
Read paper "Personalized Multimodal Generation with Large Language Models"
Fixed user post formatting
Enabled text-only posting
Friday (6/14)
Live, AI-generated responses
.... Debugging nested object arrays in MongoDB again
Week 5
Monday (6/17)
Worked on week five presentation
Another day spent debugging automated reply issues
Weekly presentations (1:30pm-3pm)
Team meeting (3pm)
Consulting Seth (lab member) about my bug (4pm)
Tuesday (6/18)
.... This bug is still plaguing me
I FIXED IT THOUGH!!!
We now have AI-generated responses to user-made posts
Started working on reply chaining
Wednesday (6/19)
Added "reply" functionality to non-user comments
Implemented AI-generated replies to replies
Debugged notifications for user replies
Fixed issue of AI-generated content being visible to subsequent users
Assigned unique user IDs to AI-generated comments
Thursday (6/20)
Completed separation of user account actions from each other
Started fixing scalability problem in feed algorithm
Created midterm presentation
Friday (6/21)
Injected more posts into database
Adjusted feed algorithm to support larger database
Cleaned posts from dataset and assigned them new timestamps
Continued working on midterm presentation
Prepped for demo by creating fake profiles and assigning them to posts
Fixed system prompting to sustain further interaction
Week 6
Monday (6/24)
10-12:30: weekly meeting
Organized script to convert sample dataset of Twitter users into database objects
Tuesday (6/26)
Cleaning and reorganizing code
Replaced band-aid fix for comments array bug
Team meeting (3-4pm)
Wednesday (6/27)
Replaced hard-coded values in Groq API prompting
Finished implementing website functionality
Wrote a script to generate posts based off of a given social media profile
Thursday (6/28)
Did research looking into cloud hosting
Incorporated tweet classification model into the post generation script
Wrote a script (.py & .js) that converts user database data into a model prediction of profile
Friday (6/29)
Generated posts and comments section for fake user profiles
Worked on week 7 presentation
Week 7
Monday (7/1)
10:30-12:30: weekly meeting
Quality control for AI profiling script
Debugging
3-4pm: team meeting
4-5pm: lab meeting
Tuesday (7/2)
Wrote in Reports schema for MongoDB
Refined AI profiling script to save a report in database
Wednesday (7/3)
Created React.js front-end for secondary profiling website
Made a main page where you can enter your Prolific ID
Implemented Express.js backend to fetch report data
Started working on slideshow
Thursday (7/4) - July 4th holiday
Friday (7/5)
Worked on reformatting front-end
Changed routing to be specify to Prolific ID as opposed to input so URL can be passed via Qualtrics
Removed email field from platform & reworked user auth system
Finished slideshow
Deployed website and had people test it
Week 8
Monday (7/8)
10:30-12: weekly meeting
Patched edge case in social media simulation
Got more testing done
3pm: team meeting
4pm: lab meeting
Tuesday (7/9)
Started restructuring report website into tabs
Integrated OpenAI into report website
More testing
Allowed social media simulation signup to take prolific ID as a parameter in URL
Removed self-commenting on synthetic content
Experimented with temperature values, settled on 0.25
Heroku crashed... don't know why
Cleaned comments to get rid of self-commenting
Wednesday (7/10)
Heroku fixed itself overnight? Weird
Fixed up prompting to be clearer and achieve more accurate results
Working on styling/CSS for profiling app
Started working on poster
Thursday (7/11)
Finishing touches on styling for report website
Restructured for production build
Deployed report website
Friday (7/12)
Poster design
Added to feed on social media simulation
Waiting for survey to be done
Week 9
Monday (7/15)
Made poster
Poster presentations (10:30am-12pm)
Initial testing
Wrote script to combine Qualtrics & Prolific csv files into a combined json file
Tuesday (7/16)
Troubleshooted & upgraded database to production tier
Added relevant database fields from websites to respective participant
Made bar graph on user sentiment for predictions
Wednesday (7/17)
Completed data collection
Reworked poster design
Examined number of interactions that participants made with platform
Thursday (7/18)
Made joint plots between number of interactions participants made and their accuracy ratings of the reports
Turned the above into a heatmap
Lots of data analysis
Friday (7/19)
Finishing up figures
Finishing up assembling poster
Week 10
Monday (7/22) - Sick day :(
Sent over poster
Started preparing deliverables
Cleaned, organized, and commented code
Delivered relevant software (Github repositories, .zip folder of data analysis) to PI
Tuesday (7/23)
Compiling deliverables
Preparing for Mid-SURE
Wednesday (7/24) - Mid-SURE
Thursday (7/24)
Finishing deliverables