Mayurkhalse/muril-sarcasm-detector

MuRIL for Sarcasm Detection (English, Hindi, & Hinglish)

This model is a fine-tuned version of Google's google/muril-base-cased specialized for Sarcasm Detection across English, Hinglish (code-mixed / transliterated Hindi), and Hindi (Devanagari script).

Model Details

• Base Architecture: Multilingual Representations for Indian Languages (MuRIL)
• Task: Binary Sequence Classification (0: Not Sarcastic, 1: Sarcastic)
• Languages: English, Hindi, Hinglish (and applicable to Indian languages)
• Regularization: Label Smoothing (0.1) & L2 Weight Decay (0.05) to ensure calibrated, non-overfit confidence scores.

How to Use in Python

You can load and use this model directly in your Python scripts with 3 lines of code:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load model and tokenizer
model_id = "YOUR_HF_USERNAME/muril-sarcasm-detector"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

# Input sentences
texts = [
    "Oh brilliant! Another meeting that could have been an email.",
    "Arey wah! 3 ghante late aakar bolte ho 'I am on time'!",
    "The package arrived safely today and was packed well."
]

# Tokenize & Predict
inputs = tokenizer(texts, padding=True, truncation=True, max_length=128, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)

for text, prob in zip(texts, probabilities):
    label = "Sarcastic" if prob[1] > 0.5 else "Not Sarcastic"
    confidence = max(prob[0], prob[1]).item()
    print(f"Text: '{text}' -> {label} ({confidence:.1%})")

Intended Use & Limitations

• Trained to identify irony, situational sarcasm, and sarcastic colloquialisms in Indian social media, workplace, and conversational contexts.