hw2-perfume-occasion-distilbert

Perfume Occasion Classifier (fine-tuned DistilBERT)

Purpose

Predicts the occasion a perfume suits — everyday, going out, formal, casual-relaxed — from a short English description of its notes and character. Built for CMU 24-679 Homework 2 to practice fine-tuning. Labels reflect one person's opinion; not a perfume recommender.

Data

• Source: ypolatog/perfume-occasion-texts (CC-BY-4.0), by classmate ypolatog.
• 100 real texts (29 everyday, 38 going out, 13 formal, 20 casual-relaxed) + 1,243 synthetic (char swap/delete, word swap, WordNet synonym).
• Split by parent text (seed 42, stratified): train 60 texts (811 incl. synthetic), validation 15 (205 incl. synthetic), test 25 real texts only. Synthetic copies never cross sets. Exact IDs in split.json (reused by the Problem 4 prompting comparison).

Preprocessing

DistilBERT uncased tokenizer, truncation at 96 tokens, dynamic padding. No other cleaning.

Training setup

• Base: distilbert-base-uncased + new 4-way classification head; full fine-tuning.
• 4 epochs, learning rate 2e-05, batch 16, weight decay 0.01, AdamW, seed 42.
• Best epoch by validation macro-F1: epoch 3 (restored at end).

Results (25 real test texts)

Confusion matrix and per-class metrics are in the training notebook. One test text ≈ 4 accuracy points, so differences of a few points are within noise.

Limitations and ethics

• Subjective single-annotator labels; many perfumes plausibly fit two occasions.
• Small classes: 13 real formal texts; synthetic copies add typos, not new meaning.
• Shortcut cues: season sentences correlate with labels (cold weather → often going out), and many texts share the opening "Prominent notes…"; the model may rely on these.
• Descriptions mention brands and "dupe of…" comparisons; the model may associate brands with occasions.

How to use

from transformers import pipeline
clf = pipeline("text-classification", model="leixiang25/hw2-perfume-occasion-distilbert")
clf("Prominent notes of iris and saffron. Elegant and refined, suits evening events.")

Compute

Google Colab, NVIDIA T4 GPU, a few minutes of training.

AI usage disclosure

Code and this card were drafted with Claude (Anthropic) and reviewed, run, and interpreted by the author.

Acknowledgments

Dataset by ypolatog (CMU 24-679, Fall 2026). Base model by Hugging Face (DistilBERT, Apache-2.0).