ccm/2026-24679-text-dataset

24-679 (Fall 2026): Haiku Perspectives

ccm/2026-24679-text-dataset

Short poems collected through the 24-679 course survey at Carnegie Mellon University. Each retained survey response contributes a human-perspective poem and an AI-or-machine-perspective poem. This dataset supports a classroom comparison of fixed embeddings, fine-tuning, and few-shot prompting.

Source and task

The preparation notebook reads 24-679-tabular-survey.csv, retains the two specified haiku columns, and reshapes paired answers into separate records. Labels come from the survey question: 0 = AI or machine perspective; 1 = human perspective. They describe the requested perspective, not verified authorship. No authorship-detection ground truth or independent expert annotation is supplied.

Preparation source: 24-679 Text Data notebook. Course: 24-679, Fall 2026, Carnegie Mellon University. Repository maintainer: the account shown above.

Fields

Field	Meaning and modeling role
haiku	Input poem text, with line breaks retained.
label	Classification target: 0 = AI/machine perspective; 1 = human perspective.
perspective, label_name	Readable target descriptions; exclude both from model inputs.
response_row	Shared survey-response key for the paired poems; grouping/provenance only.
source_id, parent_id	Unique example key and its original poem key; never input features.
augmentation, is_augmented	Method identifier and synthetic-row flag; provenance only.

Splits and original-source counts

These counts are computed from the packaged splits for this run.

Split	Original rows	Synthetic rows	Total rows
train	30	117	147
validation	6	0	6
test	8	0	8

Class counts include synthetic rows in training only.

Split	Label	Rows
train	0	73
train	1	74
validation	0	3
validation	1	3
test	0	4
test	1	4

Requested holdout fraction: 30%; test receives 50% of that holdout. Small-sample rounding changes the realized proportions. The first split uses seed 24679, and the holdout split uses seed 24680.

Split the original survey-response keys first, then carry both poems into the same partition. Every generated variant stays in training with its parent. The split is grouped by response row; it does not establish that different responses came from different people. Reuse these partitions for all model comparisons. Do not randomly re-split the augmented rows.

Augmentation and preprocessing

Character rate in this run: 30% of eligible words, capped at 3 words per poem. The corpus contains 22 distinct survey-response keys.

Each method independently starts from an original training poem; transformations are not stacked.

• Internal character swaps: swap adjacent, distinct interior letters in selected words.
• Internal character deletions: remove one interior letter from each selected word.
• Within-line word swap: exchange two distinct words in one haiku line, preserving that line's word inventory.
• WordNet synonym replacement: use NLTK's English POS tagger and the first matching WordNet sense to replace one eligible base-form noun, verb, adjective, or adverb with a single-word alternative. Proper nouns, inflected forms, pronouns, negation, and explicit human/AI/machine terms are protected or excluded.

The character methods apply the configured rate to eligible words of at least four letters, round up, and respect the per-poem cap; first and last letters remain intact. A poem with no eligible change is skipped. The WordNet first-sense rule is a heuristic, not contextual sense disambiguation.

Training method	Stored rows
internal_character_deletions	30
internal_character_swaps	30
none	30
within_line_word_swap	30
wordnet_synonym_replacement	27

Intended use and limitations

Use for teaching data preparation, augmentation audits, and perspective classification. This is a small convenience sample, not a representative corpus of human experience or machine language. Prompt wording and explicit AI/human vocabulary can make labels easy to shortcut. English-oriented matching and WordNet may handle unusual, non-English, or poetic language poorly. Augmentation can change meaning, grammar, syllable count, or perspective even when its label is copied correctly. Inspect the variants; structural assertions do not prove semantic label preservation. Report test counts alongside accuracy/F1 and compare against original-only training. Do not use these labels to identify whether a person used AI or to make decisions about students.

Privacy and licensing

The prepared schema has no dedicated name, email, or timestamp field, but free text may contain personal details or recognizable writing. Pseudonymous IDs and grouping by response row do not establish anonymity or participant consent. Review content and applicable course permissions before reuse. No license was specified in the repository metadata when this card was first added; this card does not assign one.

Load and compare

from datasets import load_dataset
ds = load_dataset("ccm/2026-24679-text-dataset")
# Train with ds["train"], choose settings with ds["validation"], then score ds["test"].

Use an account with access if repository visibility changes. For reproducible comparisons, record the dataset commit and model/environment versions. Regenerate this card with the preparation notebook after changing the data; its counts are calculated from the actual packaged splits. The YAML schema and split configuration are preserved from the upload.