This is a fine-tuned BERT-based model trained on 10,000+ Hinglish + English samples to detect human emotions from short text messages.
Unlike most emotion datasets that are purely English, this model was built to understand real Indian conversational language including Hinglish words such as:

"udas" → sad
"gussa" → angry
"mast" → joy

It powers the deployed app 👉 AI VibeCheck on Hugging Face Spaces.

📖 Model Details

Developed by: Jagrit Chaudhry
Model type: BERT for Sequence Classification
Languages: Hinglish + English (code-mixed)
Fine-tuned from: bert-base-multilingual-cased
License: MIT

🚀 Uses

Direct Use

Emotion detection from raw text (English or Hinglish).
Can process screenshots of text via OCR (in the web app).

Example:

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "Hostileic/emotion-vibecheck-model"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

inputs = tokenizer("mujhe thoda gussa aa raha hai", return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=1)
    prediction = torch.argmax(probs, dim=1).item()

print("Predicted Emotion:", model.config.id2label[prediction])

Downstream Use

Chatbots and virtual assistants that adapt to user emotions.

Emotion-aware analytics for social media or customer support.

Out-of-Scope

Long-form documents (works best on short text/snippets).

Non-Hinglish languages not present in training data.

⚠️ Bias, Risks, and Limitations

Model is biased towards Hinglish/English texting style, may underperform on formal text.

Limited coverage of rare emotions due to dataset size.

Misclassifications possible with sarcasm, irony, or mixed emotions.

📊 Training Details

Dataset: Custom synthetic + extended dataset (~10k samples, 10 emotion labels).

Training procedure: Fine-tuning bert-base-multilingual-cased with PyTorch + Hugging Face Transformers.

Hyperparameters:

Epochs: 5

Batch size: 32

Learning rate: 2e-5

Optimizer: AdamW

✅ Evaluation

Validation Accuracy: ~85%

Best performance on: Joy, Sadness, Anger

Challenging cases: Neutral and Surprise (overlaps in Hinglish texting).

⚡ Technical Specs

Architecture: BERT-base (multilingual)

Framework: PyTorch + Hugging Face Transformers

Training Hardware: NVIDIA GPU (single-GPU fine-tuning)

📌 Citation

If you use this model, please cite:

@misc{chaudhry2025emotionvibecheck,
  author = {Jagrit Chaudhry},
  title = {AI VibeCheck – Hinglish + English Emotion Detection},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Hostileic/emotion-vibecheck-model}}
}

📬 Contact

Author: Jagrit Chaudhry

Email: jagritworkchaudhry1409@gmail.com

GitHub: [Jagrit-09](https://github.com/Jagrit-09)

LinkedIn: [Jagrit Chaudhry](https://www.linkedin.com/in/jagrit-chaudhry-448690309/)