Line-Level Code Vs Text Classification Dataset

Task Definition

Binary line-level classification:

• label = 0 — natural language text
• label = 1 — programming code

The task is intentionally defined at the line granularity, rather than document or block level, to support applications such as preprocessing pipelines, mixed-content filtering, and robustness evaluation.

---

Dataset Structure

The dataset consists of three columns:

No aggressive normalization has been applied; punctuation, symbols, and formatting are preserved.

---

Data Collection and Processing

Stack Overflow Data

The Stack Overflow portion of the dataset was created by parsing HTML post bodies and explicitly separating structural elements:

• Code lines are extracted from <code> tags
• Text lines are extracted from <p> (paragraph) tags

To ensure that extracted code corresponds to actual programming content, code blocks are filtered by post tags, retaining only posts associated with the following technologies:

• <c++>, <java>, <python>, <php>, <c#>, <javascript>, <c>, <go>
• <react-native>, <laravel>, <django>, <typescript>, <node.js>, <.net-core>

Additional filtering steps:

• Lines containing package installation commands (e.g. pip) are excluded
• Lines dominated by error messages or stack traces are ignored
• Duplicate lines are removed after normalization

This filtering aims to retain clean, representative programming code while avoiding noisy or auxiliary content.

---

Twitch Data

The Twitch portion consists exclusively of natural language text and is included to introduce informal, unstructured, and ambiguous language that is difficult to distinguish from code using surface-level heuristics.

Key properties:

• No code labels are present in this subset
• Duplicate lines are removed
• Language is short, conversational, and noisy

The purpose of this source is to increase dataset difficulty and improve robustness to domain shift.

---

Dataset Statistics

Overall Size

• Total lines: 52,270
• Text (label 0): 27,330
• Code (label 1): 24,940

The dataset is approximately balanced, enabling standard classification benchmarks without mandatory resampling.

---

Label Distribution by Source

---

Source Distribution

• Stack Overflow 2020: 32,888 lines (62.9%)
• Stack Overflow 2019: 14,698 lines (28.1%)
• Twitch: 4,684 lines (9.0%)

---

Line Length Statistics (Characters)

Overall:

• Mean: 59.53
• Median: 34
• Minimum: 1
• Maximum: 9,082
• Standard deviation: 85.95

By source:

• Stack Overflow 2019: mean 88.71, median 58
• Stack Overflow 2020: mean 51.60, median 33
• Twitch: mean 23.62, median 17

---

Intended Use Cases

• Line-level code vs text classification
• Mixed-content preprocessing and filtering
• Robustness evaluation across domains
• Stylometry and authorship analysis
• Code-aware language modeling pipelines

---

Limitations

• Labels are structural, not semantic
• Code-like text outside <code> tags is labeled as text
• Logs or pseudo-code inside code blocks are labeled as code
• Line-level context across adjacent lines is not preserved

Users should account for these properties when designing downstream tasks.

---

Data Sources and Licensing

This dataset is derived from publicly available sources.
Please refer to the following datasets for original data and licensing information:

• Stack Overflow dataset: (link to be added)
• Twitch dataset: (link to be added)

Users are responsible for complying with the original licenses when redistributing or building upon this dataset.