Skye Insights

Data

Datasets from our research, hosted on Hugging Face. Both are released under CC BY-NC-ND 4.0 and cover the Skyer benchmark for large language models in pediatric emergency department triage.

The complete 55-scenario Skyer set is deliberately not released. Publishing it would let future models train on the scenario and answer pairs, which would make the benchmark meaningless. What we release instead is described below.

Skyer benchmark results

huggingface.co/datasets/SkyeInsights/skyer-results · Openly available

Per-case model predictions and reference-standard scores behind the results tables of the Skyer framework paper (DAIH workshop). It contains CTAS and ESI triage predictions, disposition decisions, and diagnostic-accuracy aggregates for 16 model versions across all 55 scenarios, each with the case number and the agreed correct answer. Decoding parameters, query dates, and per-run outputs are not yet included.

Skyer sample prompts

huggingface.co/datasets/SkyeInsights/skyer-sample-prompts · Gated, approval required

A five-scenario, canary-protected sample of the prompt set, with ground truth for initial diagnosis, triage code, and disposition encoded in base64. It is enough to see the prompt format and reproduce scoring on a small slice, without exposing the full benchmark. Access is gated on Hugging Face and granted by approval. Approved users agree not to use the data for training, fine-tuning, or redistribution.

@misc{skyerresults2026,
  author       = {{Skye Insights}},
  title        = {Skyer results: per-case model predictions and
                  reference-standard scores across 16 LLM versions},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/SkyeInsights/skyer-results}
}

@misc{skyersampleprompts2026,
  author       = {{Skye Insights}},
  title        = {Skyer sample prompts: a gated, evaluation-only subset
                  of pediatric ED scenarios},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/SkyeInsights/skyer-sample-prompts}
}

The datasets support these papers:

• Skyer benchmark (CJEM)
• Skyer framework (DAIH workshop)