Base Model: distilbert-base-uncased
Fine-Tuning: LoRA (Low-Rank Adaptation)
Architecture: "God-Mode" Rank 64, Alpha 128 (Applied across all Attention and Feed-Forward Linear layers).
Trainable Parameters: ~5.4 Million Parameters
Accuracy: 93.64% Validation Accuracy, 91.7% Real-World Edge-Case Accuracy.

The 8 Emotion Classes

The model categorizes human text into one of 8 conversational moods: 0. Sadness (Depression, exhaustion, loneliness)

Joy (Happiness, hype, energy)
Love (Affection, coziness, romance)
Anger (Frustration, aggression, sarcasm)
Fear (Anxiety, suspense, terror)
Surprise (Shock, plot twists, mind-blown)
Disgust (Apathy, revulsion, annoyance)
Calm (Relaxed, peaceful, zen)

Native MLX Deployment

This model does not rely on PyTorch or TensorFlow. The included lora_best.npz weights are completely native to Apple's Unified Memory architecture, allowing for near-instant inference times on M-series chips for chatbots.

Edge Case Performance

Unlike generic emotion models, this model has been "Surgically Grafted" to understand highly confusing conversational contexts:

“The house is completely empty and I finally have the couch to myself to just breathe” -> Correctly predicts Calm (Does not mistake "empty house" for Sadness).
“It’s way too quiet in this house and I keep hearing noises” -> Correctly predicts Fear (Does not mistake "quiet" for Calm).
“Wow I am speechless I did not see that plot twist coming” -> Correctly predicts Surprise (Does not mistake "speechless" for Disgust/Joy).

DiyRex/emo-movies-lora

EmoMovies LoRA Engine 🤖🎬

Model Details

The 8 Emotion Classes

Native MLX Deployment

Edge Case Performance

DiyRex/
emo-movies-lora