birgermoell/swedish-health-source-triage

Swedish Health Source Triage

This is a small custom text-classification dataset for an Information Retrieval assignment about embeddings. The task is to classify short health-information texts by the public source family they resemble:

• —1177.se: patient-facing healthcare guidance
• —socialstyrelsen.se: national authority reports, guidelines, and statistics
• —lakemedelsverket.se: medicine and medical-product regulation

Intended use

The dataset is intentionally compact and readable. It is useful for teaching:

• —source classification
• —collection routing
• —embedding-based baselines
• —train/test evaluation
• —model cards and demo Spaces

It is not a medical system and must not be used for health advice.

Data creation

The examples were manually written as domain-shaped teaching data. They do not copy official texts. The labels and source families are based on public-facing roles of the three Swedish health-information sources.

The design follows the assignment idea of creating a custom dataset from a specific domain, similar in spirit to synthetic or manually curated datasets used for privacy-preserving NLP prototyping.

Columns

• —source: class label
• —source_type: human-readable source family
• —title: short document title
• —body: short document body
• —url: representative source homepage

Splits

The repository ships a single JSONL file. Training scripts create a stratified train/test split with a fixed random seed. When uploaded to Hugging Face, the file can be used as the train split or manually split into train and test files.

Limitations

This is a teaching dataset, not a benchmark. Strong scores on this data only show that the methods learn the constructed source signal. A serious deployment would require larger samples, real source documents, deduplication, topic balancing, Swedish-language text, and external validation.