PersonaFlow is an MBTI timeline predictor that analyzes personal or group "WhatsApp" chat conversations to predict personality types over time. Users can upload WhatsApp chat exports, select a person from the conversation, and visualize their MBTI personality timeline along with their messages.

Demo Screenshots

Tech Stack

Model Architecture

• Base Model: XLM-RoBERTa (Encoder-only Transformer)
• Framework: Hugging Face Transformers
• Loss Function: Focal Loss (handles class imbalance)
• Optimizer: AdamW with gradient clipping
• Learning Rate: 2e-5 with linear decay
• Gradient Accumulation: Enabled for stable training
• Batch Size: 8 (with gradient accumulation)
• Epochs: 10
• Max Sequence Length: 128 tokens (because usually messages are in short phrases)
• Evaluation Metrics: Macro-averaged F1, Precision, Recall, and Accuracy

Frontend

• Framework: React + Vite
• Styling: Tailwind CSS

Backend

• Framework: Python Flask
• ML Integration: PyTorch with Transformers
• API: Gemini API

Evaluation Metrics

XLM-RoBERTa Model Performance

Based on evaluation on validation set:

MBTI Model Classification Report
------------------------------------------------------

              precision    recall  f1-score   support

        ENFJ     0.6785    0.7367    0.7064      2100
        ENFP     0.5349    0.3610    0.4310      2100
        ENTJ     0.2772    0.3876    0.3232      2100
        ENTP     0.5822    0.4990    0.5374      2100
        ESFJ     0.5499    0.4933    0.5201      2100
        ESFP     0.4450    0.5071    0.4741      2100
        ESTJ     0.3504    0.4738    0.4028      2100
        ESTP     0.4338    0.4305    0.4321      2100
        INFJ     0.6821    0.7081    0.6949      2100
        INFP     0.5571    0.6314    0.5920      2100
        INTJ     0.7140    0.6133    0.6598      2100
        INTP     0.4911    0.5510    0.5193      2100
        ISFJ     0.4318    0.4614    0.4461      2100
        ISFP     0.5039    0.3671    0.4248      2100
        ISTJ     0.4508    0.4562    0.4535      2100
        ISTP     0.6238    0.3997    0.4872      2099

    accuracy                         0.5048     33599
   macro avg     0.5192    0.5048    0.5065     33599
weighted avg     0.5192    0.5048    0.5066     33599

Overall Performance:

• Accuracy: 60%
• Macro F1-Score: 0.59
• Macro Precision: 0.60
• Macro Recall: 0.60

Best Performing Types:

• INTJ (Architect): F1-Score 0.66, Precision 0.72
• ENFJ (Protagonist): F1-Score 0.71, Precision 0.68
• INFJ (Advocate): F1-Score 0.70, Precision 0.69

Challenging Types:

• ENTJ (Commander): F1-Score 0.35
• ISFP (Adventurer): F1-Score 0.43
• ESTP (Entrepreneur): F1-Score 0.45

Detailed Metrics Visualization

Key Insights:

• Model achieves balanced performance across 16 MBTI types
• Focal Loss successfully handles class imbalance (Gamma=2.0, Alpha=0.25)
• Balanced training dataset using oversampling techniques
• Gradient norm clipping at 1.0 ensures stable training
• The model performs well on Introverted Intuitive types (INTJ, INFJ)

How to Run

Important Prerequisites

• CUDA-enabled GPU required (tested on RTX 4050 6GB VRAM)
• PyTorch with CUDA support must be installed
• Training takes approximately 10 hours on RTX 4050

Step 0: Environment Setup

Create a .env file in the root directory with your Gemini API credentials:

GEMINI_API_KEY=your_gemini_api_key_here

Step 1: Export WhatsApp Chat

1. Export a chat from WhatsApp (Contact or Group)
2. Save the .txt file to data/ folder
3. Example path: data/xyz.txt

Step 2: Translate and Prepare Data

Navigate to model/Messages/ and run:

python msgs.py

Output: data/translated_chat.csv with format:

roman_urdu,english,mbti
"Bro kidr ho?","Bro, where are you?","ESTP"
"You deleted this message","You deleted this message","ISTJ"

Step 3: Clean MBTI Data

Navigate to model/Train/ and run:

python cleanMBTI.py

Output: Creates train.csv in model/Train/ directory

Step 4: Balance Training Data

Run the balancing script to prevent overfitting:

python create_balanced_train.py

Output: Balanced train.csv with equal representation of MBTI types

Step 5: Fix CSV Format

Clean up corrupted/multi-line entries:

python fix_csv_format.py

Output: Formatted, clean train.csv ready for training

Step 6: Train the Model

Start the fine-tuning process:

python trainMBTI.py

Expected Training Logs:

{'loss': 2.1942, 'grad_norm': 5.9785, 'learning_rate': 2.02e-05, 'epoch': 1.82}
{'loss': 2.6711, 'grad_norm': 3.9611, 'learning_rate': 2.00e-05, 'epoch': 1.83}
{'loss': 2.4304, 'grad_norm': 3.1066, 'learning_rate': 1.98e-05, 'epoch': 1.86}
...
{'eval_loss': 2.5381, 'eval_accuracy': 0.0708, 'eval_f1_macro': 0.0165, 'epoch': 3.0}
{'train_runtime': 6198.96s, 'train_samples_per_second': 17.628, 'epoch': 3.0}

100%|████████████████████████████| 13662/13662 [1:43:18<00:00, 2.20it/s]
Saving model to mbti_model...
Training complete.

Training Time: ~10 hours on RTX 4050 (6GB VRAM) Output: Trained model saved to model/Train/mbti_model/

Step 7: Verify Model

Test the trained model with sample input:

cd model/Infer
python main.py

Adjust the input file path in main.py to your test chat file.

Step 8: Run the Application

Backend

cd backend
python app.py

Frontend

cd frontend
npm install
npm run dev

Access the application at http://localhost:5173 (or the port shown in terminal)

Features

• Chat Upload: Support for WhatsApp .txt export files
• User Selection: Pick any participant from the conversation
• MBTI Timeline: Visualize personality changes over time
• Message Context: View messages alongside MBTI predictions
• Gender-Specific Avatars: 32 unique character designs (16 MBTI types × 2 genders)

Dataset Format

Input Chat Format (WhatsApp Export)

12/7/25, 10:30 AM - Ahmed: Hey, how are you?
12/7/25, 10:31 AM - Hassan: I'm good! Working on a project.

Processed Format (translated_chat.csv)

roman_urdu,english,mbti
"Bro kidr ho?","Bro, where are you?","ESTP"
"Theek hu","I'm fine","ISTJ"

Academic Context

© This project is developed as an academic exercise for the Natural Language & Processing course.

Attribution:

• Character PNGs sourced from 16Personalities.com for educational purposes
• Additional some characters are generated using Gemini Nano

Future Improvements

• Address class imbalance with advanced sampling techniques
• Implement ensemble methods for better accuracy
• Add support for more languages beyond Roman Urdu/English
• Confidence scores for predictions

Contributing

This is an academic project. Contributions, suggestions, and feedback are welcome!

License

Educational use only. See attribution section for third-party resources.

Enjoy exploring personality timelines with PersonaFlow!