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.
Vran is an experimental RESEARCH model with NO medical intended purpose — not a medical device (EU MDR), not for diagnosis, treatment or clinical use. Research and non-commercial use only. You accept sole responsibility for regulatory/legal compliance (EU AI Act, GDPR).
Log in or Sign Up to review the conditions and access this model content.
Vran-27B-SLO-BioMed-Research — LoRA adapter
Trained by MediaAtlas — LLM fine-tuning on your own data, trained in the EU, weights delivered. Pricing · All our models
Na kratko: Raziskovalni slovenski biomedicinski jezikovni model (27B), adapter LoRA. Samo za raziskave in nekomercialno rabo; ni medicinski pripomoček. Na cjvt/slovenian-llm-eval je izenačen z GaMS3-12B-Instruct po log-likelihood in pred njim pri neposrednem vprašanju (0,852 proti 0,773). Sprejme tudi slike.
Vision / slike: the adapter was trained on text only; after merging, the base model's vision encoder still handles image input. Image understanding was not evaluated in Slovenian.
The LoRA adapter (PEFT, rank 128 / alpha 256, lora_target: all) for
texdata/Vran-27B-SLO-BioMed-Research
— a Slovenian biomedical research language model with native tool-calling.
⛔ Not a medical device, not for clinical use. Research / non-commercial only.
Base model — apply to the right one
This adapter was trained on texdata/Qwen3.6-27B-slo-med-mt
(a Qwen3.6-27B SL-medical MT base), not plain Qwen3.6-27B. Applying it to a different
base will not reproduce Vran.
from peft import AutoPeftModelForCausalLM
model = AutoPeftModelForCausalLM.from_pretrained("texdata/Vran-27B-SLO-BioMed-Research-LoRA",
trust_remote_code=True) # pulls the base automatically
For a ready-to-use merged model use texdata/Vran-27B-SLO-BioMed-Research
(safetensors) or its GGUF build.
Evaluation (summary)
Slovenian-LLM-Eval (cjvt/slovenian-llm-eval), 0-shot, n = 500 examples/task (3,500 total), scored
against GaMS3-12B-Instruct with the same code on the same examples:
| protocol | Vran 27B | GaMS3-12B-Instruct |
|---|---|---|
| asked through the chat template (letter answer) | 0.852 | 0.773 |
log-likelihood acc_norm (lm-eval convention) |
0.670 | 0.663 |
Tied under the log-likelihood convention, ahead when both models are asked the question directly. Internal medical eval (130 questions, 13 domains, multi-judge): pravilnost 4.99 / popolnost 4.97. Per-task tables, confidence intervals and caveats are on the main model card.
Lineage
Qwen3.6-27B → uncensored "heretic" variant → texdata/Qwen3.6-27B-slo-med-mt → Vran. The chain
passes through a model with safety alignment removed, so Vran carries no upstream guardrails —
see the main card.
Training
Full SFT: ~161k examples (138k Slovenian medical + 19k tool-calls + identity + retention), 2 epochs. Eval (multi-judge): pravilnost 4.99 / popolnost 4.97. See the merged model card for details, disclaimers, and lineage.
- Downloads last month
- 2
Model tree for texdata/Vran-27B-SLO-BioMed-Research-LoRA
Base model
Qwen/Qwen3.6-27B