ashish5193/sarcasm_model

Hindi/Hinglish Sarcasm Detector

Overview

This model is a fine-tuned version of:

cardiffnlp/twitter-roberta-base-sentiment-latest

for Hindi/Hinglish sarcasm detection.

Dataset

Dataset used:

Huggmachas/Sarcasm_dataset

Dataset size:

• Total samples: 5250
• Non-Sarcastic (0): 4746
• Sarcastic (1): 504

Labels

• 0 = Not Sarcastic
• 1 = Sarcastic

Training Details

• Base Model: RoBERTa
• Epochs: 3
• Learning Rate: 2e-5
• Batch Size: 16
• Framework: Hugging Face Transformers

Results

Validation Loss:

• Best Validation Loss: 0.1256

Example

Input:

Maiden shabash bas ab aisay he khelo -.- #Sarcasm

Output:

Sarcastic

Input:

Aaj mausam bahut accha hai

Output:

Not Sarcastic

Limitations

The dataset is highly imbalanced:

• Non-Sarcastic: 4746
• Sarcastic: 504

During evaluation, the model showed a tendency to rely on Twitter-style sarcasm markers such as:

• #Irony
• #Sarcasm

The model may not generalize well to everyday conversational sarcasm.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="ashish5193/sarcasm_model"
)

result = classifier(
    "Maiden shabash bas ab aisay he khelo -.- #