Dataset and Labels

Dataset and Labels

The script embeds 30 manually written support requests. The labels are:

Class	Index	Examples of topics
billing	0	Charges, invoices, plans, refunds
technical	1	API errors, app crashes, failed integrations
account	2	Account settings, access, ownership
feature_request	3	Requested features and integrations

train_test_split(test_size=0.2, random_state=42, stratify=labels) produces 24 training and 6 validation rows. Because six validation rows are too few for dependable conclusions, this is a pipeline demonstration, not a benchmark.

Preprocessing

AutoTokenizer.from_pretrained('distilbert-base-uncased') tokenizes text with truncation=True, max_length=128 and padding=False; DataCollatorWithPadding pads dynamically per batch.

Improving the dataset

Collect a substantially larger, representative, consented, de-identified dataset; split by time or customer as appropriate to avoid leakage; add label audits and a held-out test split; evaluate multiple seeds. Do not upload private tickets without authorization.