danijmg/smart-campus-tickets

Smart Campus IT Support Tickets (synthetic, teaching)

Short IT support tickets from a fictional university campus, each labelled with one of four categories. Built for a three-session teaching course on applied AI and practical machine learning (University of Malaga lecturer, Kauno Kolegija, October 2026).

Origin

• Synthetic. Every ticket was written by Claude (Anthropic) and reviewed by the lecturer on 2026-10-03. No real tickets, people, buildings or organisations.
• The review covered an automatic report (class counts, seeded defects, near-duplicates, ambiguous tickets), all flagged tickets and 10 random tickets per class; debatable tickets were rewritten or replaced before freezing.

Intended use

Teaching only: data cleaning, train/test discipline, a classical text classifier (TF-IDF + logistic regression), and comparison with language models on the same frozen test set. Results are illustrative: synthetic text is cleaner and more regular than real tickets, so high scores here say little about real-world performance. Do not use it to benchmark models for publication or production.

Files

File	Content	Rows
tickets_v1.csv	All tickets, including the seeded defects	410
test_ids.csv	Frozen test ticket_id (20 per label)	80
dev_ids.csv	TRAIN ids for prompt work (5 per label)	20
cpu_subset_ids.csv	TEST ids for the CPU path (5 per label)	20

The three id files have one column, ticket_id.

Columns (tickets_v1.csv)

Column	Type	Values
ticket_id	text	T0001, T0002, ...
created_at	ISO date	YYYY-MM-DD
building	text	fictional campus building
channel	text	email, portal, phone
priority	text	low, medium, high
text	text	the ticket, in English
label	text	account, iot, network, printer

Columns are listed in file order. 400 valid tickets, 100 per label (balanced).

Seeded defects

Defects are injected on purpose so that students practise cleaning.

Defect	Count
Malformed labels (wrong case or trailing spaces)	8
Empty texts (load as NaN in pandas)	4
Exact duplicate rows	6
Raw rows -> clean rows	410 -> 400

410 rows = 400 valid tickets + 6 duplicates + 4 empty texts.

Cleaning rule

Apply in this order, then split:

1. Strip and lower-case label.
2. Drop rows whose text is missing, empty or only spaces.
3. Drop exact duplicate rows (all columns identical, after step 1).

The result must have 400 rows, 100 per label, unique ticket_id.

import pandas as pd

DATA_BASE_URL = ("https://huggingface.co/datasets/"
                 "danijmg/smart-campus-tickets/resolve/main")
df = pd.read_csv(f"{DATA_BASE_URL}/tickets_v1.csv")
df["label"] = df["label"].str.strip().str.lower()
df = df[df["text"].fillna("").str.strip() != ""].drop_duplicates()
assert len(df) == 400

Use the /resolve/main/ address; the /tree/main address is the web page.

Frozen split

• Clean rows sorted by ticket_id, then scikit-learn train_test_split with test_size=80, stratify=label, random_state=42: 320 train / 80 test, 20 per label in test.
• dev_ids.csv: 5 per label from TRAIN, seed 42, drawn only from tickets already clean in the raw file.
• cpu_subset_ids.csv: 5 per label from TEST, seed 42.
• Ids were chosen by seed, never by looking at model results. Test tickets must never be used to tune prompts, rules, labels or parameters.
• Produced by the notebook AIML_freeze_split.ipynb; the uploaded id files, not a re-run, are the reference.

Versioning

• Frozen file names never change and frozen files are never edited.
• New content means a new version file (for example tickets_v2.csv) with its own id files (for example test_ids_v2.csv); test_ids.csv, dev_ids.csv and cpu_subset_ids.csv always refer to tickets_v1.csv.
• Course material pins the repo commit used in class.

License

CC BY 4.0. Attribution: "Smart Campus IT Support Tickets (synthetic), Claude-authored, reviewed by Daniel-Jesus Munoz (University of Malaga), 2026."