| --- |
| license: other |
| license_name: research-noncommercial |
| language: |
| - sl |
| tags: |
| - medical |
| - biomedical |
| - slovenian |
| - evaluation |
| - benchmark |
| - research |
| pretty_name: Slovenian Medical Eval (Vran) |
| task_categories: |
| - question-answering |
| - text-generation |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # Slovenian Medical Eval — med-slo-eval |
|
|
| Evaluation suite for **Slovenian biomedical language models** (used to score the |
| [Vran](https://huggingface.co/texdata/Vran-27B-SLO-BioMed-Research) research model). |
| Two parts: |
|
|
| > **Research / educational use only. Not for clinical use.** These are test items for |
| > evaluating language models, not medical guidance. |
|
|
| ## 1. Open-ended benchmark (`benchmark/`) |
|
|
| Held-out, hand-curated **open-ended clinical questions in Slovenian**, scored by an |
| LLM judge on **pravilnost** (correctness) and **popolnost** (completeness), 1–5. |
|
|
| | file | items | format | |
| |---|---|---| |
| | `benchmark/edge_by_domain.jsonl` | 130 | `{"id", "domain", "q"}` — 13 domains × 10 (kardiologija, nevrologija, pediatrija, farmakologija, urgentna_medicina, …) | |
| | `benchmark/hard_questions.jsonl` | 50 | `{"id", "category", "q"}` — hard clinical vignettes | |
|
|
| No reference answers are shipped (judged by model, not exact-match) — keeping the |
| questions **out of training sets** avoids benchmark contamination. |
|
|
| ## 2. Held-out test split (`test/`) |
|
|
| `test/test.jsonl` — **7,144** unified-schema SFT records **with reference answers** |
| (the held-out test split of the `med-slo-sft` corpus). Same schema as |
| [`texdata/med-slo-sft`](https://huggingface.co/datasets/texdata/med-slo-sft): |
| `messages` (system/user/assistant), `domain`, `metadata`, etc. Use for held-out |
| loss/quality evaluation. Split by content-based `dedup_cluster` → no leakage from train. |
|
|
| ## Usage |
|
|
| Answer the benchmark questions with your model (Slovenian, thinking optional), then judge |
| each answer for correctness/completeness with a strong LLM judge (1–5). Report per-domain |
| and overall means. For the test split, score generated answers against the references or |
| compute held-out perplexity. |
|
|
| ## License |
|
|
| **Research and non-commercial use only.** The benchmark is authored; the test split is a |
| derivative of open medical datasets (translated) — respect the most restrictive upstream |
| license. You are responsible for applicable law (EU AI Act, GDPR). |
|
|
