# Setup MCP
Use the Bright Data web MCP skill to setup mcp with pro mode. Here is my token: <your_brightdata_token>


# Get reviews
Scraped Play Store reviews from 12 AI/education apps using Bright Data Web MCP:
https://play.google.com/store/apps/details?id=com.openai.chatgpt
https://play.google.com/store/apps/details?id=com.google.android.apps.bard
https://play.google.com/store/apps/details?id=com.microsoft.copilot
https://play.google.com/store/apps/details?id=ai.perplexity.app.android
https://play.google.com/store/apps/details?id=com.anthropic.claude
https://play.google.com/store/apps/details?id=com.microblink.photomath
https://play.google.com/store/apps/details?id=co.brainly
https://play.google.com/store/apps/details?id=com.quizlet.quizletandroid
https://play.google.com/store/apps/details?id=com.duolingo
https://play.google.com/store/apps/details?id=com.cliffweitzman.speechify2
https://play.google.com/store/apps/details?id=com.codeway.wonder
https://play.google.com/store/apps/details?id=com.starryai
Create a new folder "reviews/raw/" and save the actual comments in a txt file under the folder. Save 12 text files.


# Process reviews
Process all text files in the "reviews/raw/" folder. Create csv with columns:
1. app_name (from filename)
2. review_text (the actual review comment within "")
3. category (one of: Bug/Issue, Feature Request, Praise, Complaint)


# Fine-tuning
Start a new fine-tuning experiment using SFT for text classification task.
- Maintain a report for experiment. 
- Use "distilbert/distilbert-base-uncased" model to fine-tune on the "reviews/reviews_analysis.csv" dataset.
- Use `review_text` as input and `category` as output.
- Use local GPU for training.
- Use tensorboard with local dashboard for monitoring.