REALMResearch project · S-ITP-26-11
Development of Reasoning-Based Small Language Models for Oncology Data
We build compact, transparent language models that reason step by step over clinical oncology data — designed to run securely within healthcare institutions.
Focus areas
Reasoning
Models that explain their conclusions through verifiable, step-by-step reasoning.
Small & efficient
Compact models deployable on local hospital infrastructure, keeping data private.
Oncology data
Adapted to clinical records, pathology reports and cancer-specific terminology, including Lithuanian.
First results
The project's first public outputs: a Lithuanian biomedical decision dataset and two small models fine-tuned on it.
MEDLINE-S1-LT
Lithuanian translation of MEDLINE-S1: typed questions over PubMed abstracts — study design, main disease area, and yes/no questions about humans, animals, adverse effects and drug therapy. Labels come from NLM MeSH indexing, not from a language model. This release adds the Lithuanian title and abstract.
- Licence
- CC BY 4.0
- Splits
- train · dev · calib · test
- Translated with
- NLLB-200 1.3B
laya-multilingual-lt-med
Fine-tune of laya-multilingual on MEDLINE-S1-LT. One forward pass returns a probability for every answer option; the model does not generate text.
- Decision accuracy
- 0.924
- Record accuracy
- 0.628
- Calibration error
- 0.004
medgemma-1.5-4b-it-decitions-lt-med
LoRA adapter with a decision head on MedGemma 1.5 4B IT, trained on MEDLINE-S1-LT. Scores every question on an abstract in a single pass.
- Decision accuracy
- 0.946
- Record accuracy
- 0.723
- Calibration error
- 0.005
Model scores are measured on the MEDLINE-S1-LT held-out test split — 4,857 abstracts and 29,654 typed decisions.