yennik16/ny-pa-nj-text

NY / PA / NJ — Short Regional Texts

Purpose

This dataset was built for Carnegie Mellon University course 24679 HW1. It is intended as an exercise in collecting and augmenting text data, which can be binarily classified. Ultimately, the data is separated into ~200 character texts that either describe New Jersey or another state. This could be used to train a simple model that predicts whether or not a text is talking about New Jersey. It can also provide the ability to compare how the various augmentation methods used affect accuracy and results over the original data.

Composition

100 passages, each roughly 200 characters (min 154, median 214, max 318), covering food, infrastructure, law, transport, geography, industry, and culture.

column	description
text	the passage
is_nj	binary target — 1 if the passage describes New Jersey
state	secondary three-way label: NJ / PA / NY
char_count, word_count	length metadata
source_id, parent_id	identity and lineage
severity, augmentation, is_augmented	provenance of synthetic rows

Label balance: 50 NJ / 50 non-NJ (exactly 50/50 by construction). The non-NJ half is 25 PA and 25 NY.

Collection

Every sentence was written and assigned to a state by me. The only outside source used for data collection was the occasional google search to fact check something I was hazy on, in order to preserve the accuracy of the data (such as the details of Punxsatawney Phil to reach 200 chars).

Preprocessing

Whitespace trimmed; state codes uppercased; is_nj derived from state. No text was edited, filtered, deduplicated, or truncated — there were no duplicates and no missing values.

Augmentation

Noise injection only. No paraphrasing, no synonym substitution, no back-translation, no generative model. Applied to the 70 training passages only: 14 variants per passage, plus one extra light variant for 20 randomly chosen passages, 1000 synthetic rows total.

Operations: character deletion, adjacent-character transposition, QWERTY-neighbour substitution, and character duplication; dropout of function words; adjacent word swaps; formatting corruption (dropped apostrophes and commas, missing space after a period, lost capitalisation); and sentence shuffling.

tier	copies	char rate	dropout	swaps	formatting	shuffle
light	4 (+1 for 20 passages)	2%	5%	0	no	no
moderate	4	5%	10%	1	yes	no
heavy	3	10%	15%	2	yes	no
shuffled	3	2%	5%	0	no	yes

Label preservation

The label is copied from the parent and never recomputed, so it cannot drift. The substantive risk is that a corruption destroys the evidence for the label rather than the label itself — the signal lives in a small set of proper nouns (Shoprite, Wawa, Sheetz, yinz, jughandle), and deleting one leaves a passage that could describe any of the three states.

Mitigation: 70 proper nouns are identified from the training passages (internal capitals, or capitalisation in non-sentence-initial position, minus function words) and are never deleted and receive at most one character-level typo, so they stay fuzzy-matchable. Only function words are eligible for dropout. Sentence shuffling is safe because the passages are lists of independent statements whose order carries no meaning.

Every synthetic row is verified distinct from its parent and from every other row.

Splits

• original — all 100 passages I wrote.
• train — 70 of those passages (~70%; 35 NJ / 17 PA / 18 NY), followed by the 1000 synthetic passages derived from them, for 1070 rows.
• test — 15 passages (~15%; 8 NJ / 4 PA / 3 NY), unaugmented.
• validation — 15 passages (~15%; 7 NJ / 4 PA / 4 NY), unaugmented.

The real passages in train, test, and validation partition original, stratified on state, so every passage appears in exactly one of them. Every synthetic row descends from a train passage via parent_id, so no test or validation passage has near-identical copies in train and the held-out splits can be used for evaluation as-is. Filter train on is_augmented to train on the real passages only.

Intended use and limits

This data is intended as an exercise for 24679 HW1 to build familiarity with collecting and augmenting text based data. In its current state, it is intended to be used to build a basic classifier that tries to identify whether or not a text is talking about New Jersey.

• The data is limited in the fact that there are only 100 short texts, all written by me. This means that the potential details for a model to rely on when classifying are limited to what was included based on my own recollection. Additionally, my own writing style and phrasing habits may influence potential classifiers broader accuracy.

• The initial data I wrote consists of 50 texts about NJ, 25 texts about PA, and 25 texts about NY (the latter two being grouped as not NJ). It is therefore limited in the states it actually covers and has a more detailed coverage of NJ. This was due to the fact that I found it more difficult to come up with 50 texts about PA, supplementing with texts about NY as I am less familiar with each.

• The actual texts and classifications are based on subjective observations and are not verified.

Ethical notes

No human subjects, no personal data, no third-party or scraped content, no copyrighted material. Passages describe places and regional customs, not individuals. Some content touches on regional stereotypes (driving habits, accents, sports loyalties); while these are stereotypes they are in no way meant to be malicious or offensive. They do not represent assumptions of any one individual but instead broad regional generalizations, most of which are commonly known/accepted affectionately.

License

CC BY 4.0.

AI usage disclosure

This dataset and the code to generate it were vibecoded as outlined by the CMU 24679 guidelines. With that being said, all of the original data was written/collected by me manually, with the occasional google search to verify a text I had already written so as not to introduce incorrect data. The observations themselves are entirely my own.

In generating all code and the augmented dataset, Claude was used by first providing it with the example from class, the assignment description, and my generated dataset. I specifically instructed it to not immediately generate new code in order to discuss details first. I prompted it with my proposed augmentation methods (drop random words, swap letters, drop letters, etc.) based on the methods described in class. It recommended the label preservation and proper noun retention methods and handled them as described above. I also had Claude generate the intial draft of this Dataset Card, mostly for the purpose of handling formatting. It has since been heavily revised and rewritten by me, including this section.

To accomodate the assignment change to the splits required as quickly as possible, the .ipynb was fed back into claude with the following prompt "modify this dataset such that the splits are an original, a training split consisting of about 70% of the data, a test split consisting of about 15%, and a validation split consisting of the last 15% rather than the current splits. The 1000 augmented data points should be made entirely of the 70% training split using the same methods as previous. If these are not sufficient to hit 1000 augmented samples, let me know before continuing so we can discuss next steps. The augmented samples should be appended to the end of the training split so it consists of 70% of the original data plus 1000 augmented samples. Upload it to the same huggingface found here: https://colab.research.google.com/drive/14F5KPQ9r71MaYUb36SXuwzXZ94KOSS6S?usp=sharing Change comments and the dataset card as necessary to reflect the new splits in the same style. Do not edit the ai usage disclosures".

It did pause to discuss further augmentation methods to meet the required samples with the new shortened training set. At this point I opted to just use the same methods to create more iterations for simplicity.

It was then prompted a few additional times to resolve errors resulting from the dataset card, and the new output reviewed.