lbakar/health-log-extraction-dataset
To create this dataset, we sampled demographic seeds from an occupation and age-range table from the Labor Force Statistics[1]. For each occupation, the pipeline randomly selected an age range, sampled an age within that range, and assigned a gender from a fixed set of options. These demographic seeds were used to prompt an LLM to generate structured personas containing a name, description, medications or supplements, general mood, and possible health conditions or injuries. We then used… See the full description on the dataset page: https://huggingface.co/datasets/lbakar/health-log-extraction-dataset.
To create this dataset, we sampled demographic seeds from an occupation and age-range table from the Labor Force Statistics<sup>[[1]](https://www.bls.gov/cps/cpsaat11b.htm)</sup>. For each occupation, the pipeline randomly selected an age range, sampled an age within that range, and assigned a gender from a fixed set of options. These demographic seeds were used to prompt an LLM to generate structured personas containing a name, description, medications or supplements, general mood, and possible health conditions or injuries.
We then used each generated persona to synthesize labeled health-log entries. A Jinja prompt template incorporated the persona attributes, the target category set, and few-shot examples of both single-label and multi-label logs. Gemini generated structured JSON outputs conforming to a predefined schema, where each example consisted of a message, modality, and one or more category labels. The resulting JSONL dataset was used to train a multi-label classifier over six categories: Activities, Mood, Symptoms, Treatment, Food, and NONE.
Subsequently, we ran NuExtract3 on each example of the dataset to get the extraction ground truth.
References
[1] U.S. Bureau of Labor Statistics, Employed persons by detailed occupation and age, 2025. Available at: https://www.bls.gov/cps/cpsaat11b.htm