You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access you acknowledge that Vran is an experimental RESEARCH language model with NO medical intended purpose. It is not a medical device (EU MDR), is not CE/FDA-cleared, and is not for diagnosis, treatment, triage, prescribing, clinical decision-making, or patient care. You agree to use it only for research and non-commercial purposes, not to deploy it as a clinical or "health assistant", and you accept sole responsibility for compliance with applicable law (EU AI Act, GDPR, EU MDR). Provided AS IS, without warranty; the authors accept no liability.
Log in or Sign Up to review the conditions and access this model content.
Vran — Slovenian Biomedical Research Language Model (27B)
Trained by MediaAtlas — LLM fine-tuning on your own data, trained in the EU, weights delivered. Pricing · All our models
Vran is an experimental research language model for Slovenian medical / biomedical text, built by supervised fine-tuning of Qwen3.6-27B (dense) on a full Slovenian medical corpus. It explains medical and scientific concepts in fluent, formal Slovenian for research and educational purposes, and supports native tool / function calling (Qwen3.6 XML format) for agentic research use.
⛔ Not a medical device. Not a clinical assistant.
Vran has no medical intended purpose. It is not a medical device under EU MDR 2017/745, is not CE-marked or FDA-cleared, and is not intended for diagnosis, treatment, triage, monitoring, prescribing, clinical decision-making, or patient care. It is released for research and non-commercial use only. It does not replace a qualified clinician.
Intended purpose (regulatory framing)
The intended purpose is research in natural-language processing and medical/biomedical NLP, and education. Any use that informs an actual clinical or health decision — for a patient, by a clinician, or by a member of the public — is out of scope and not permitted. Deployers who repurpose the model for a medical intended purpose assume all resulting regulatory obligations.
⚠️ Disclaimers
- Not medical advice. Outputs are generated text and may be incorrect, incomplete, outdated, or fabricated (hallucinated); they must not be relied on for any health decision.
- No clinician substitute. Always consult a licensed clinician.
- No warranty, no liability. Provided "AS IS". The authors accept no liability for any harm arising from use. The user assumes all responsibility and risk.
- Compliance is the user's. You are responsible for the EU AI Act, GDPR, MDR and any other applicable law in your jurisdiction.
Evaluation
Internal Slovenian medical eval suite (130 questions, 13 clinical domains), multi-judge (two independent LLM judges, averaged to de-bias), 1–5, run with the research system prompt:
| metric | score (1–5) |
|---|---|
| pravilnost (correctness) | 4.99 |
| popolnost (completeness) | 4.97 |
Per-domain (pravilnost / popolnost):
| domena | pravilnost | popolnost |
|---|---|---|
| endokrinologija | 5.00 | 5.00 |
| farmakologija | 5.00 | 4.95 |
| gastroenterologija | 5.00 | 4.95 |
| ginekologija/porodništvo | 5.00 | 5.00 |
| hematologija | 5.00 | 4.95 |
| infektologija | 4.95 | 4.90 |
| kardiologija | 4.95 | 4.90 |
| nefrologija | 5.00 | 5.00 |
| nevrologija | 5.00 | 5.00 |
| pediatrija | 5.00 | 5.00 |
| pulmologija | 5.00 | 4.95 |
| toksikologija | 5.00 | 4.95 |
| urgentna medicina | 5.00 | 5.00 |
Tool / function calling
Vran emits native Qwen3.6 XML tool calls when tools are provided, e.g.:
<tool_call><function=get_weather><parameter=city>Ljubljana</parameter></function></tool_call>
Trained on Hermes / APIGen-MT / ToolACE tool-call data alongside the medical corpus, so agentic/function-calling ability is retained without degrading Slovenian medical quality.
General Slovenian ability — head-to-head with GaMS3-12B-Instruct
Medical fine-tuning did not cost general Slovenian competence. cjvt/slovenian-llm-eval, 0-shot,
n = 500 examples per task (3,500 total). Both models were scored by the same code on the same
examples — leaderboard numbers are not comparable across harnesses, so the reference model was
re-run locally rather than quoted. Two protocols are reported, because they measure different things:
(a) Answered as a question — the task is rendered as a lettered multiple-choice question, put through each model's chat template, and the answer letter is parsed (greedy, no reasoning block; no answer was unparseable for either model):
| task | n | Vran 27B | GaMS3-12B-Instruct |
|---|---|---|---|
| arc_easy | 500 | 0.984 [0.97–0.99] | 0.952 [0.93–0.97] |
| arc_challenge | 500 | 0.950 [0.93–0.97] | 0.878 [0.85–0.90] |
| hellaswag | 500 | 0.616 [0.57–0.66] | 0.560 [0.52–0.60] |
| piqa | 500 | 0.844 [0.81–0.87] | 0.734 [0.69–0.77] |
| openbookqa | 500 | 0.886 [0.86–0.91] | 0.794 [0.76–0.83] |
| winogrande | 500 | 0.784 [0.75–0.82] | 0.628 [0.58–0.67] |
| boolq | 500 | 0.898 [0.87–0.92] | 0.866 [0.83–0.89] |
| AVERAGE | 3500 | 0.852 | 0.773 |
Vran leads on all seven tasks (+7.9 points on average, outside the ±~1.2-point interval at this n). Caveat worth stating: Vran's SFT mix contains a lot of lettered medical multiple-choice, so part of this gap is format familiarity, not only knowledge.
(b) Scored by log-likelihood — the lm-eval-harness convention (acc_norm, char-normalized) on
raw text, no chat template, which is what most published Slovenian numbers use:
| task | n | Vran 27B | GaMS3-12B-Instruct |
|---|---|---|---|
| arc_easy | 500 | 0.760 | 0.756 |
| arc_challenge | 500 | 0.566 | 0.562 |
| hellaswag | 500 | 0.592 | 0.572 |
| piqa | 500 | 0.702 | 0.704 |
| openbookqa | 500 | 0.470 | 0.488 |
| winogrande | 500 | 0.718 | 0.714 |
| boolq | 500 | 0.884 | 0.846 |
| AVERAGE | 3500 | 0.670 | 0.663 |
Under this protocol the two models are tied (0.7 points, well inside noise). Log-likelihood scoring does not let an instruction-tuned model use its chat format, so it understates both models — by ~18 points here — and compresses the difference between them. [·] = 95% Wilson interval.
Summary claim: Vran is on par with GaMS3-12B-Instruct under the standard log-likelihood protocol, and ahead of it when both models are asked the question directly. Caveats: these tasks are machine-translated multiple-choice benchmarks — they measure Slovenian knowledge and reasoning, not generation quality or medical correctness; 500/task is a subsample of the 20,465-example test set; neither protocol used the reasoning block.
(Internal benchmarks; not a substitute for clinical validation.)
Model & identity
Presents itself as "Vran", a Slovenian research language model — explicitly not a doctor and not a medical device. Multimodal (text + vision) Qwen3.6 architecture; MTP tensors preserved for speculative decoding.
Lineage note: the Qwen3.6-27B in that chain is reached through an uncensored intermediate — see Training data & lineage below.
Training data & lineage
- Base (direct):
texdata/Qwen3.6-27B-slo-med-mt— our Slovenian medical EN⇄SL translation fine-tune, not plain Qwen3.6-27B. Its own base isllmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, an uncensored ("heretic") variant of Qwen3.6-27B (dense) with safety alignment removed — chosen so the translation stage would not refuse or soften clinical content. Full chain:Qwen3.6-27B → heretic (uncensored) → slo-med-mt (SL medical MT) → Vran (SL medical SFT + tool calls). - Safety consequence, stated plainly: Vran inherits no upstream safety alignment. Its SFT re-establishes the research identity and the "not a clinician" framing, but refusal behaviour is not a hardened guardrail and must not be relied on. Deployers are responsible for their own filtering.
- Slovenian medical corpus — our curated "golden" corpus, machine-translated from open
English medical datasets (MedMCQA, MedQA-style, PubMedQA, MedReason, MedQuAD) with a strict
term-preservation gate. Released separately so others can reproduce/extend:
texdata/med-slo-sft. - Slovenian conversational data from
cjvt/GaMS-Nemotron-Chat(retention, to keep general Slovenian fluency). Vran is not the GaMS model; only some conversational data is shared. - Tool-calling data — Hermes / APIGen-MT / ToolACE (English), for native function calling.
- Trained as a full SFT (
161k examples: 138k medical + 19k tool-calls + identity + retention); medical stays dominant (86%) so tool-calls don't dilute Slovenian medical quality.
License
Derivative of Qwen3.6-27B; the upstream Qwen license applies. As a derivative work, this release adds a research-only, non-commercial restriction. No commercial use.
Citation
@misc{vran_27b_slo_medical_research,
title = {Vran: a Slovenian biomedical research language model (27B)},
author = {texdata},
year = {2026},
note = {Research artifact, non-commercial. Derivative of Qwen3.6-27B. Not a medical device.}
}
- Downloads last month
- -