ccm/2026-24679-tabular-dataset

24-679 (Fall 2026): Music Listening Survey

ccm/2026-24679-tabular-dataset

Course-survey responses about music listening, plus explicitly marked synthetic training variants. The classroom regression task predicts weekly listening hours from two count features and five categorical preferences. Stored haiku answers provide source context and are excluded from this task's predictors.

Source and task

The preparation notebook reads 24-679-tabular-survey.csv, removes the form timestamp, assigns source IDs in retained row order, and checks the selected fields. Original answers are self-reports from the 24-679 course survey at Carnegie Mellon University. Listening time must be finite and between 0 and 168 hours per week; song and playlist counts must be finite, nonnegative integers. These checks establish valid domains, not the truth of a respondent's estimate.

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

Fields

The machine-readable feature metadata at the top of this card preserves every exact column name and storage type.

Splits and original-source counts

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

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.

Randomly split original response rows before augmentation, then retain both holdouts unchanged. Both parents of every Mixup row must come from the original training partition. IDs are assigned from row order; changing or reordering the source CSV can change regenerated IDs and splits. The split does not identify repeat submissions from one person. Keep these prepared boundaries fixed for downstream model comparisons.

Augmentation and preprocessing

This run requests 3 copies per method and original training row before filtering.

Every draw starts from original training rows. The copy-count setting requests the same number of draws per method and source; unchanged rows and repeated feature/target combinations for a given primary parent are removed.

• —Additive numeric jitter: perturb only song and playlist counts with Gaussian noise. Each standard deviation is max(5% of that training column's IQR, 1 count); clip below zero and round to integers.
• —Multiplicative numeric scaling: independently multiply those two counts by factors in 0.90–1.10 and round to nonnegative integers.
• —Regression Mixup: select two distinct original training parents. Apply the same primary weight in 0.60–0.90 to both count predictors and the listening-hours target. Round counts only. Copy the primary parent's entire categorical profile and other stored context; record both parents and the weight.
• —Categorical perturbation: change one categorical predictor to a different training-observed choice. Sharing and attendance move one adjacent step on the ordered Never–Rarely–Sometimes–Often–Very Often scale; other choices use a different observed category.

Jitter, scaling, and categorical perturbation keep the target unchanged. Original and single-parent rows repeat the primary key in second_parent_id and use mix_weight=1.0. SMOTE and CTGAN are not used.

Intended use and limitations

Use for teaching mixed-type data contracts, provenance, augmentation, and small-sample regression. This convenience sample does not represent the general population. Self-reports may contain recall error, outliers, or different interpretations of the questions. A domain-valid synthetic row may describe an implausible listener: Mixup assumes a useful interpolation, while categorical perturbation assumes the original target remains appropriate. More rows do not add independent respondents or establish better generalization. Compare original-only and augmented training with the same validation/test sets; report MAE and RMSE in hours/week and a median-of-original-training baseline. Do not use this classroom sample for individual profiling or consequential decisions.

Privacy and licensing

The timestamp is removed, but listening preferences and retained free-text poems may be identifying in combination. Source IDs do not establish anonymity or participant consent. The raw CSV remains a separate source artifact. Review content and applicable course permissions before reuse. No license was specified in 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-tabular-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.