from transformers import pipeline

clf = pipeline("text-classification", model="kash620/fashion-classifier")

clf("Levi's 511 Slim Fit Stretch Jeans")
# [{'label': 'Bottoms', 'score': 0.99}]

Training Data

~157,000 real product titles scraped from fashion e-commerce stores, balanced to ~15,000 samples per class. Stratified 90/10 train/test split.

Training Details

Hyperparameter	Value
Base model	distilbert-base-cased
Max sequence length	64
Epochs	5
Batch size	32 (train) / 64 (eval)
Learning rate	3e-5
Warmup ratio	0.1
Loss	Cross-entropy with inverse-frequency class weights
Mixed precision	FP16

Cased tokenizer was chosen to preserve brand names and acronyms. Class weights handle residual imbalance after downsampling.

Performance

Evaluated on a stratified 10% holdout (~15,700 samples).

Metric	Score
Accuracy	~97%
Weighted F1	~0.97

All 12 classes exceed 0.94 F1. The model handles ambiguous titles well (e.g. "Oversized Utility Shirt Jacket" → Outerwear, not Tops).

Performance

Evaluated on a stratified 10% holdout (~15,700 samples).

Metric	Score
Accuracy	97%
Weighted F1	~0.97

Citation

If you use this model, please also cite the original DistilBERT paper:

bibtex

@article{sanh2019distilbert,
  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
  author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
  journal={arXiv preprint arXiv:1910.01108},
  year={2019}
}

Limitations

English-only
Fashion product titles only — not suitable for general text classification
Some inherently ambiguous products (e.g. shackets, co-ord sets) may be inconsistently classified
May reflect brand distribution biases present in the scraped data