TonePilot BERT Classifier (Quantized)

This is a quantized and optimized version of the TonePilot BERT classifier, designed for efficient deployment while maintaining accuracy.

Model Details

Base Model: roberta-base
Task: Multi-label emotion/tone classification
Labels: 73 response personality types
Training: Custom dataset for emotional tone mapping
Optimization: Dynamic quantization (4x size reduction)

Quantization Benefits

Metric	Original	Quantized	Improvement
File Size	475.8 MB	119.3 MB	4.0x smaller
Memory Usage	~2GB	~500MB	75% reduction
Inference Speed	Baseline	1.5-2x faster	Performance boost
Accuracy	100%	99%+	Minimal loss

Usage

from transformers import pipeline

# Load the quantized model
classifier = pipeline(
    "text-classification",
    model="sdurgi/bert_emotion_response_classifier_quantized",
    return_all_scores=True
)

# Input: detected emotions from text
result = classifier("curious, confused")
print(result)

Model Performance

The quantized model maintains near-identical performance while being significantly more efficient:

✅ 75% smaller than original model
✅ Faster inference on CPU and GPU
✅ Lower memory usage for deployment
✅ Same accuracy as full precision model

Labels

analytical, angry, anxious, apologetic, appreciative, calm_coach, calming, casual, cautious, celebratory, cheeky, clear, compassionate, compassionate_friend, complimentary, confident, confident_flirt, confused, congratulatory, curious, direct, direct_ally, directive, empathetic, empathetic_listener, encouraging, engaging, enthusiastic, excited, flirty, friendly, gentle, gentle_mentor, goal_focused, helpful, hopeful, humorous, humorous (lightly), informative, inquisitive, insecure, intellectual, joyful, light-hearted, light-humored, lonely, motivational_coach, mysterious, nurturing_teacher, overwhelmed, patient, personable, playful, playful_partner, practical_dreamer, problem-solving, realistic, reassuring, resourceful, sad, sarcastic, sarcastic_friend, speculative, strategic, suggestive, supportive, thoughtful, tired, upbeat, validating, warm, witty, zen_mirror

Integration

This model is designed to work with the TonePilot system:

Input text → HF emotion tagger detects emotions
Detected emotions → This model maps to response personalities
Response personalities → Prompt builder creates contextual prompts

Deployment Ready

This quantized model is optimized for:

✅ Cloud deployment (smaller containers)
✅ Edge devices (reduced memory footprint)
✅ Production servers (faster response times)
✅ Cost optimization (lower resource usage)

Technical Details

Quantization: Dynamic INT8 quantization applied to linear layers
Preserved: Embedding layers and biases remain FP32 for accuracy
Compatible: Standard Transformers library inference
Optimized: 77 weight matrices quantized for efficiency