ElderRyan/psycial

MBTI Personality Classifier - Multi-Task Model

Production-grade MBTI (Myers-Briggs Type Indicator) personality classification model implemented in Rust with GPU acceleration.

Model Description

This model predicts MBTI personality type from text using a multi-task learning approach with four independent binary classifiers for each MBTI dimension:

• E/I: Extraversion vs Introversion
• S/N: Sensing vs Intuition
• T/F: Thinking vs Feeling
• J/P: Judging vs Perceiving

Architecture

• Input Features: 5384 dimensions
  • 5000 TF-IDF features (top words from vocabulary)
  • 384 BERT embeddings (sentence-transformers/all-MiniLM-L6-v2)
• Network: 4-layer deep MLP
  • Architecture: 5384 → [1024, 768, 512, 256] → 4×2 outputs
  • Dropout: 0.5 (disabled during inference)
  • Optimizer: Adam (lr=0.001)
• Training: Per-dimension epochs with weighted loss
  • E/I: 30 epochs, weight=1.2
  • S/N: 30 epochs, weight=1.0
  • T/F: 25 epochs, weight=1.0
  • J/P: 30 epochs, weight=1.3

Performance

Overall Metrics

Metric	Value
Test Accuracy	49.80%
vs Random Baseline (6.25%)	8.0x better
vs TF-IDF Baseline (21.73%)	+129.2% improvement
Training Time	~50 seconds (GPU)

Per-Dimension Accuracy

Dimension	Accuracy	Samples
E/I	80.58%	1398/1735
S/N	87.15%	1512/1735
T/F	81.90%	1421/1735
J/P	75.33%	1307/1735

Training Data

• Dataset: MBTI Kaggle Dataset (8675 samples)
• Split: 80% train (6940), 20% test (1735)
• Classes: 16 MBTI types (INTJ, ENFP, etc.)
• Class Distribution:
  • I: 77%, E: 23% (imbalanced)
  • N: 86%, S: 14% (highly imbalanced)
  • F: 54%, T: 46% (balanced)
  • J: 60%, P: 40% (moderately imbalanced)

⚠️ Important Note

This model does NOT support Hugging Face Inference API because it's a custom Rust/PyTorch implementation, not a standard transformers model.

To use this model, you have two options:

Option 1: Use the Web Interface (Easiest)

Visit https://psycial.0xbase.ai for instant MBTI predictions - no setup required!

Option 2: Run Locally (Full Control)

Usage

Requirements

• Rust 1.70+
• PyTorch/libtorch (for tch-rs bindings)
• CUDA (optional, for GPU acceleration)

Installation

git clone https://github.com/RyanKung/psycial
cd psycial

# Set up environment
conda create -n psycial python=3.10
conda activate psycial
conda install pytorch

# Build
export LIBTORCH_USE_PYTORCH=1
export LIBTORCH_BYPASS_VERSION_CHECK=1
cargo build --release

Download Model from Hugging Face

from huggingface_hub import hf_hub_download

# Download model files
mlp_weights = hf_hub_download(
    repo_id="ElderRyan/psycial",
    filename="mlp_weights_multitask.pt"
)

vectorizer = hf_hub_download(
    repo_id="ElderRyan/psycial",
    filename="tfidf_vectorizer_multitask.json"
)

# Copy to models directory
import shutil
shutil.copy(mlp_weights, "models/mlp_weights_multitask.pt")
shutil.copy(vectorizer, "models/tfidf_vectorizer_multitask.json")

Inference

# Download model files first (see above)

# Predict single text
./target/release/psycial hybrid predict "I love solving complex problems and thinking deeply about abstract concepts."

Programmatic Usage (Rust)

use psycial::hybrid::predict::predict_single;

// Load model and predict
predict_single("Your text here")?;

Quick Test

# Download inference example
wget https://huggingface.co/ElderRyan/polyjuice/raw/main/inference_example.py

# Run it
python inference_example.py

Model Files

This model repository includes:

1. mlp_weights_multitask.pt (27MB) - Neural network weights
2. tfidf_vectorizer_multitask.json (213KB) - TF-IDF vocabulary and IDF weights
3. feature_normalizer.json (5KB) - Feature normalization parameters (optional)
4. feature_selector.json (1KB) - Pearson feature selector indices (optional)

Limitations

1. Moderate Accuracy: 49.80% is significantly better than random/baseline but still has room for improvement
2. Class Imbalance: Model may favor majority classes (I, N, F, J)
3. Data Bias: Trained on online forum posts, may not generalize to all text types
4. Language: English only
5. MBTI Validity: MBTI itself has limited scientific validity

Ethical Considerations

• MBTI is not scientifically validated for hiring/clinical decisions
• Predictions should be used for entertainment/research only
• Be aware of class imbalances and potential biases
• Do not use for discriminatory purposes

Technical Details

Framework

• Language: Rust
• ML Framework: tch-rs (PyTorch bindings)
• BERT Model: sentence-transformers/all-MiniLM-L6-v2
• Device: CUDA-capable GPU (falls back to CPU)

Key Innovations

1. Dropout Bug Fix: Discovered and fixed critical bug where dropout remained active during inference, causing ~5% accuracy loss
2. Per-Dimension Optimization: Different epochs and loss weights for each MBTI dimension
3. Multi-Task Learning: Four independent binary classifiers instead of 16-way classification

Citation

If you use this model, please cite:

@software{psycial_mbti_2025,
  title = {Psycial: Multi-Task MBTI Personality Classifier},
  author = {ElderRyan},
  year = {2025},
  url = {https://huggingface.co/ElderRyan/psycial},
  github = {https://github.com/RyanKung/psycial}
}

License

GNU General Public License v3.0 (GPLv3) - See LICENSE file for details

Acknowledgments

• rust-bert: Guillaume BE for BERT Rust implementation
• Dataset: MBTI Kaggle Dataset
• sentence-transformers: all-MiniLM-L6-v2 model