ssg1/study-spots-text

Study Spots — Good or Not (Text)

100 short descriptions of places on CMU and UC Davis campuses, each labelled by whether the spot is good for studying.

Property	Value
Splits	train (1151), validation (15), test (15)
Target	label — 1 good for studying, 0 not
Balance (all originals)	50 good / 50 not good
Length	200 characters on average

Purpose

Binary text classification where the label is not stated in the text. A passage mentions outlets, noise, seating or foot traffic; the verdict has to be inferred from those details rather than read off a sentiment word.

Exploratory analysis

Lengths overlap almost completely across the two classes, so length carries no signal. Some words lean one way — quiet and outlets toward the positive class, loud and crowded toward the negative — but many passages reach their verdict through specifics instead.

Composition

Column	Type	Description
source_id	string	stable key
parent_id	string	source passage; equals source_id for an unaugmented row
campus	string	CMU or UC Davis
text	string	the passage
label	int	target — 1 good for studying, 0 not
augmentation	string	none or the method that produced the row
is_augmented	bool	whether the row is derived

Collection

Each passage describes one place on a campus I know well and runs to roughly 200 characters. I wrote them from my own experience of both campuses, then used an AI assistant to proofread and even out the style across the set. The judgments, places and details are mine.

label = 1 means I would go there to study. The verdict never appears literally in the text.

The set spans libraries, cafes, lounges, corridors and outdoor spots on both campuses, and sentence structure varies deliberately — some passages open with the place, some with the verdict, some recount a specific visit.

Augmentation

Up to 18 variants are attempted per training passage, cycling through five methods with seed 24679; 1081 survive the unchanged-copy check. Validation and test passages are never used as parents:

Method	What it changes
synonym_substitution	one or two words swapped for close equivalents
sentence_reordering	sentences presented in a different order
contraction_toggle	it is ↔ it's and similar
hedge_insertion	a hedging phrase added at the start or end
typographical_noise	one internal character swapped or deleted in a single word

Label-preserving by construction. The verdict in each passage rests on concrete details, and none of these methods removes a detail or reverses its sense: synonym pairs are curated so loud never becomes quiet, a reordered passage describes the same place, and a typo leaves the word recognisable. Labels are inherited from the parent and never recomputed.

Every child is compared against its parent and discarded if identical, so unchanged copies do not appear. 179 were discarded on that basis.

Splits and intended use

Split	Passages	Contents
train	1151	70 training originals + 1081 variants
validation	15	Held-out originals, unaugmented
test	15	Held-out originals, unaugmented

Originals were split about 70/15/15, stratified on the label, before augmentation, so no validation or test passage has a descendant in train. Use the splits as given.

from datasets import load_dataset

ds = load_dataset("ssg1/study-spots-text")
train, validation, test = ds["train"], ds["validation"], ds["test"]

Suitable for fine-tuning a compact transformer such as DistilBERT, or for few-shot prompting experiments.

Limitations

• One author, two campuses. Every judgment is mine, so the label reflects how I study — quiet, seated, laptop and outlet — rather than any general standard.
• 100 original passages is small, so validation holds 15 and test holds 15. A single misclassification moves test accuracy by roughly 7 points.
• Augmentation adds surface variation, not new places, so training-set performance is optimistic. The typographical noise method in particular produces text that is slightly less natural than the originals.
• Named places may act as shortcuts. A model can learn that passages mentioning a specific library tend to be positive, which is memorisation rather than comprehension.
• Campus-specific vocabulary means the dataset will not transfer to other universities without adaptation.

Ethical notes

The passages describe public campus spaces, not people. No individual is named or described, and nothing here identifies a person. The judgments are personal opinions about places, not factual claims about the institutions.

License

MIT.

AI usage disclosure

• Passages — written by the author from personal experience of both campuses. Used AI to speed up corrections, formatting and make the style consistent across the set. The judgments, places and details are the author's.
• Synthetic samples — deterministic Python string transforms (synonym substitution, reordering, contraction toggling, hedging, typographical noise), reproducible from seed 24679. No generative model was used to produce them.
• Notebook and card — Claude helped structure the notebook and draft documentation. Topic, labels, augmentation methods and the review of every passage were done by the author.