Hari5115/hinglish-retail-intent-classifier

Hinglish Retail Intent Classifier

Fine-tuned version of google/muril-base-cased for classifying Indian ecommerce customer support messages (Hinglish) into 13 intent categories.

Performance

• Test accuracy: 97.6%
• Macro F1: 97.6%

Label	Precision	Recall	F1	Support
cancel_order	1.00	1.00	1.00	22
damaged_item	1.00	1.00	1.00	23
delivery_complaint	0.83	0.87	0.85	23
discount_query	1.00	1.00	1.00	23
exchange_product	1.00	1.00	1.00	21
other	1.00	1.00	1.00	19
payment_issue	1.00	1.00	1.00	23
product_availability	1.00	1.00	1.00	23
product_information	1.00	1.00	1.00	21
refund_status	1.00	1.00	1.00	23
return_request	1.00	1.00	1.00	23
track_order	0.95	0.82	0.88	22
wrong_item	0.91	1.00	0.95	21

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Hari5115/hinglish-retail-intent-classifier"
)

result = classifier("mera order kab aayega?")
print(result)
# [{'label': 'track_order', 'score': 0.96}]

Intents supported

Label	Description
track_order	Customer asking where their order is or when it arrives
cancel_order	Customer wants to cancel a placed order
exchange_product	Customer received wrong size/colour, wants a swap
refund_status	Customer asking about refund timeline or whether it was processed
delivery_complaint	Late delivery, not delivered, delivery agent issue
damaged_item	Product arrived broken, damaged, or defective
wrong_item	Completely wrong product was delivered
return_request	Customer wants to return a product
payment_issue	Money deducted but order not placed, double charge
product_availability	Asking if an item is in stock or available in a size/colour
product_information	Asking about product details, material, dimensions
discount_query	Promo code not working, coupon issues, asking for discount
other	Anything that does not fit the above

Training data

Trained on Hari5115/hinglish-retail-intent-dataset

Training config

• Base model: google/muril-base-cased
• Epochs: 5
• Learning rate: 2e-5
• Batch size: 16 (train), 32 (eval)
• Max sequence length: 128
• Best model selected by macro F1 on validation set

Limitations

• Synthetic data — may not capture all real-world linguistic variation
• Primarily Hindi-English mixing; does not cover Tamil-English, Telugu-English, or other Indian language combinations
• Skewed toward common intents; rare edge cases are underrepresented