leixiang25/24679-hw1-text-garments

24-679 HW1 (Fall 2026): Garment Descriptions

leixiang25/24679-hw1-text-garments

100 original garment descriptions written by Lei Xiang for 24-679 Homework 1 at Carnegie Mellon University, plus explicitly marked synthetic training variants. The classification task predicts the garment type (0 top, 1 bottom, 2 outerwear, 3 dress, 4 footwear) from a product-listing style description of about 200 characters.

Source and task

Every description is the author's own writing, produced while looking at retail product photos and without reading the retailer's copy. No brand names, product names, or prices appear. The author trained in fashion design, so descriptions use working garment vocabulary (silhouette, fabric, construction details, closure). Labels were assigned at writing time from the garment being described.

Fields

Splits

Stratified random split of original descriptions before augmentation: holdout 30%, half to test. Seed 24679 for the first split, 24680 for the holdout split. Validation and test contain only original text.

Class counts include synthetic rows in training only.

Augmentation and preprocessing

This run requests 5 draws per method and training original, each with a distinct seed; unchanged copies and repeated texts from the same parent are removed. Every draw starts from the original text; methods are never stacked.

• —Internal character swaps: swap two adjacent interior letters in 30% of eligible words (4+ letters), rounded up, at most 3 words; first and last letters preserved.
• —Internal character deletions: delete one interior letter in the same number of words.
• —Within-sentence word swap: exchange two distinct words inside one sentence; punctuation and spacing stay in place.
• —WordNet synonym replacement: replace one noun, verb, adjective, or adverb with a lemma from its first WordNet sense. Garment nouns (jacket, trousers, dress, boot, and so on) are protected and never replaced because they carry the label; function words are protected as well.

Intended use and limitations

Teaching dataset for text augmentation auditing and small-sample classification. Descriptions follow a deliberately consistent listing structure, so a classifier can rely on the garment noun alone; the task is easy by design. Character noise can make some words hard to read, and word swaps can break grammar; both are inherited-label assumptions to inspect, not claims that the variant is natural text. Synthetic rows do not add independent garments. Report per-class F1 with the validation and test splits.

Privacy and licensing

Original fictional product descriptions by the author; no personal data, brands, or copied retailer text. No license is assigned by this card.

Load

from datasets import load_dataset
ds = load_dataset("leixiang25/24679-hw1-text-garments")