itsjhuang/watsonx-docs-document-type

Watsonx Docs Document Type Classification

This dataset is a balanced binary document-level classification subset derived from ibm-research/watsonxDocsQA.

Task

Classify IBM Watsonx documentation pages by their dominant user-facing purpose:

• conceptual: documents primarily used to understand or look up information.
• how-to: documents primarily used to complete a procedure or fix a problem.

Splits

Split	conceptual	how-to	Total
train	140	140	280
validation	30	30	60
test	30	30	60

Fields

• doc_id: original document ID from the source dataset.
• url: source documentation URL.
• title: documentation page title.
• text: model input text, constructed as title + "\n" + first 800 words of document. The title is preserved in full; the document body is truncated to keep inputs manageable for embedding-based classifiers.
• label: string label, either conceptual or how-to.
• label_id: numeric label ID, where conceptual = 0 and how-to = 1.
• split: dataset split.

Usage

from datasets import load_dataset

data_files = {
    "train": "train.csv",
    "validation": "validation.csv",
    "test": "test.csv",
}

dataset = load_dataset("csv", data_files=data_files)

Curation Notes

IBM technical documentation has traditionally been structured around DITA (Darwin Information Typing Architecture), which classifies documents into four types: task, concept, reference, and troubleshooting. This dataset adapts that taxonomy into two classes: conceptual merges concept and reference (both primarily information-seeking); how-to merges task and troubleshooting (both action- or fix-oriented). The binary schema was chosen because troubleshooting was too rare to form a reliable standalone class, and reference and concept were difficult to separate consistently.

Annotation followed a semi-automatic process. Labelling rules were first defined based on IBM Writing Style guidelines, then applied by a heuristic script to generate candidate labels. Each candidate was assigned a confidence tier: title_high (strong title signal), body_medium (body-text signal only, no strong title match), or default_low (no strong signal in either title or body). All tiers except body_medium how-to rows were manually reviewed. The body_medium how-to subset (333 rows) was left unreviewed because the remaining manually checked data was sufficient to construct a balanced 400-example dataset; retaining unreviewed borderline rows would have introduced noise without benefit.

Rows marked X during manual review were removed because the source document was incomplete or too ambiguous to label reliably. Rows marked ? were interpreted as belonging to the opposite binary class.

The final subset contains 400 examples, sampled with random seed 42 after manual correction and filtering.

This dataset is derived from ibm-research/watsonxDocsQA, which is licensed under Apache 2.0. This dataset inherits the same license.