User message
→ language detection
→ translate to English if non-English
→ emotion classifier
→ intent classifier
→ RAG retrieval
→ final response generation

Important note

The model was trained on an English emotion dataset. For Arabic, French, Spanish, German, or other languages, translate the user input to English first, then run the classifier.

Metrics

Test Accuracy: 0.921
Test Macro-F1: 0.8897

Example

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo_id = "MariamMohsen112/mental-health-emotion-roberta"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)
model.eval()

text = "I feel anxious and I cannot sleep"

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)

probs = torch.softmax(outputs.logits, dim=-1)[0]
pred_id = int(torch.argmax(probs).item())

print(model.config.id2label[pred_id], float(probs[pred_id]))