| --- |
| license: cc-by-4.0 |
| base_model: distilbert-base-uncased |
| library_name: transformers |
| tags: |
| - text-classification |
| - fashion |
| - product-categorization |
| - indian-ecommerce |
| language: |
| - en |
| --- |
| # Fashion Category Classifier |
| DistilBERT fine-tuned to classify Indian fashion product titles into 11 categories. |
|
|
| ## Credits |
| Built on [DistilBERT](https://huggingface.co/distilbert/distilbert-base-uncased) by Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf (Hugging Face), licensed under [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0). |
|
|
| ## Labels |
| `Tops` `Bottoms` `Dresses` `Outerwear` `Footwear` `Bags` `Jewellery` `Ethnicwear` `Activewear` `Innerwear` `Headwear` |
|
|
| ## Usage |
| ```python |
| from transformers import pipeline |
| clf = pipeline("text-classification", model="roaringguts/DistilBERT") |
| clf("Nike Dry Fit Running Tshirt") |
| # [{'label': 'Activewear', 'score': 0.99}] |
| # batch |
| clf(["Lavie Women Tote Bag", "Malabar Gold Plated Necklace", "Clarks Oxford Shoes"]) |
| ``` |
|
|
| ## Training |
| - **Base model:** `distilbert-base-uncased` |
| - **Dataset:** ~63k real product titles from Myntra + synthetic samples generated for underrepresented categories |
| - **Split:** 80/10/10 train/val/test, stratified |
| - **Epochs:** 5 (early stopping, patience 2) |
| - **Batch size:** 64 |
| - **Learning rate:** 3e-5 |
| - **Precision:** fp16 |
|
|
| ## Evaluation |
| Evaluated on a held-out stratified test set. |
|
|
| | Metric | Score | |
| |--------|-------| |
| | Accuracy | **99.27%** | |
|
|
| ## Limitations |
| - Trained on Indian e-commerce titles — may underperform on Western brand naming conventions |
| - Ethnicwear and Innerwear have some overlap (e.g. sports bras vs regular bras) |
| - Low sample count for Bags and Headwear in original data, partially filled with synthetic titles |
|
|
| ## License |
| This model is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The base DistilBERT weights it derives from are licensed under Apache 2.0. |
