0408happyfeet/markers-tip-binary

Markers Grip Dataset — Binary Tip Size

Purpose

Educational dataset for binary tabular classification predicting marker tip style (fine vs bold) from physical/container features.

Dataset composition

• —Original split: 30 unique, real‑world measurements (no duplicates).
• —Augmented split: 300 rows via label‑preserving Gaussian jitter of numeric features.
• —Features (5): container_length_mm (int), grip_diameter_mm (float), length_to_diameter (float), ink_family (str), color (str).
• —Target (binary): label ∈ {0: fine, 1: bold}. Mapping: bold iff original tip_size_mm ≥ 1.0.

Class balance (original): fine=9, bold=21.

Data collection

Physical measurements of consumer markers (length and grip diameter) plus categorical attributes (ink family, color). No web‑scraped or synthetic sources in the original split.

Preprocessing & augmentation

• —Engineering: length_to_diameter = container_length_mm / grip_diameter_mm.
• —Augmentation: Gaussian jitter with σ = 0.05×std per numeric feature; values clipped to the original min/max; integer fields rounded back to int; labels unchanged. Rationale: introduces small, measurement‑scale perturbations without changing semantics (label).

Labels

• —label: 0 = fine, 1 = bold. Derived deterministically from original tip_size_mm (≥1.0 → bold).

Splits

Two splits in the Hub repo: original, augmented.

Intended use & limits

• —Intended for classroom exercises on tabular pipelines (EDA, preprocessing, training, evaluation).
• —Not suitable for high‑stakes decisions. Tiny sample size; narrow domain.

Ethical notes

• —Consumer product measurements; no personal data.
• —Avoid over‑interpreting fairness metrics due to small N and categorical sparsity.

License

• —CC BY 4.0. Provide attribution if you reuse.

AI usage disclosure

• —No generative models used to create data or labels.
• —This README and augmentation code were authored with assistance from an LLM (ChatGPT); human verified.

EDA (original split)

Summary stats (numeric): see dataset preview.

[image] [image]