sarcasm_xlmr

This model is a fine-tuned version of xlm-roberta-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1423
• Accuracy: 0.9498
• F1: 0.9546

Model description

Hinglish Sarcasm Detection Model (XLM-RoBERTa Base — Fine-Tuned)

Version: 1.0 Author: Amaan Shaikh (amaan00z) Language: Hinglish Labels: 0 = not_sarcastic 1 = sarcastic

---

✨ Model Summary

This is a Hinglish Sarcasm Classifier trained on a rich combination of:

✅ MUStARD Hinglish Dialogues ✅ Swami et al. Hinglish Twitter dataset ✅ 4,000 generated Gen-Z + political + meme sarcasm samples ✅ 2,500 real non-sarcastic Hinglish social-media samples ✅ Reddit & general Hinglish sarcasm templates ✅ Cleaned + deduplicated final dataset: 9,594 samples

Model backbone: XLM-RoBERTa Base

---

📘 Labels

ID Label

0 not_sarcastic 1 sarcastic

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
0.2244	1.0	505	0.1979	0.9220	0.9291
0.2034	2.0	1010	0.1418	0.9387	0.9448
0.1759	3.0	1515	0.1423	0.9498	0.9546
0.1009	4.0	2020	0.2033	0.9465	0.9509

Framework versions

• Transformers 4.57.1
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.22.1