FEEL WISE

FeelWiseEmotion

DEVELOPED by SANJITH RJ & SHAIK BAVIKADI AL HAFEEZ

FeelWiseEmotion is a transformer-based emotion recognition model tailored for real-world applications that require nuanced understanding of human emotions. From detecting joy in social media posts to identifying distress in mental health messages, FeelWiseEmotion delivers semantically rich emotion classifications optimized for Large Language Models (LLMs) and downstream NLP tasks.

What Makes FeelWiseEmotion Stand Out?

Feature	Description
Transformer Encoder Backbone	Built using a custom multi-head self-attention encoder with `AddNorm`, `FeedForward`, and dropout.
Emotion-Aware Output	Classifies text into six universal emotions: joy, sadness, anger, fear, surprise, and disgust.
LLM-Optimized Representations	Mean pooling generates compact, interpretable vectors ideal for prompt engineering and fine-tuning.
Fully Configurable	Easily adjust depth, embedding dimensions, number of heads, FFN size, and vocab size via config.
Hugging Face Compatibility	Integrated with `AutoModel` and `AutoTokenizer` for frictionless usage in production pipelines.

Architecture Highlights

Multi-Head Self-Attention Encoder with tunable parameters.
Custom FeedForward Layers for deeper representation learning.
Mean Pooling + Linear Classifier for stable emotion class logits.
Modular & Extensible Design compatible with Hugging Face transformers library.

Installation & Usage

pip install transformers torch

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "sanjithrj/FeelWiseEmotion"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "I'm feeling really overwhelmed and anxious today."
inputs = tokenizer(text, return_tensors="pt")
logits = model(**inputs).logits
pred = torch.argmax(logits).item()
print(f"Predicted Emotion: {model.config.id2label[pred]}")

Example Output

logits = model(**inputs).logits
# Output: tensor([[-0.18, 2.73, 0.41, 1.12, 0.09, -0.45]])
# → Predicted class index: 1 → Emotion: "Sadness"

FeelWiseConfig (Custom)

{
  "model_type": "FeelWiseEmotion",
  "n_layers": 1,
  "d_model": 256,
  "n_head": 8,
  "d_ff": 1024,
  "max_len": 500,
  "input_vocab_size": 50000,
  "dropout": 0.1,
  "num_classes": 6
}

Best Suited For

Chatbots with emotional intelligence
Customer Feedback Analysis
Mental Health Monitoring Tools
Social Media Sentiment Apps
Survey and Review Classification

Limitations

Trained on English-only datasets.
Emotion boundaries can be subjective—best results with short, personal text.
Mean pooling may underperform on longer documents with multiple sentiment shifts.

Dataset

Trained on the dair-ai/emotion dataset:

A popular benchmark for emotion classification covering multiple real-world expressions.

This model brings robust emotion understanding to the NLP pipeline with a lightweight yet extensible architecture—ideal for developers, researchers, and product teams building emotionally aware AI systems.

Would you like a badge for transformers, datasets, or model cards added as well?

Downloads last month: 32

Safetensors

Model size

13.6M params

Tensor type

F32

sanjithrj
/

FeelWise