Home

The training data we prepared
for MINA's new version

Preparing a clinical AI model starts with data, long before training. We are sharing the scale of the training data we prepared for MINA's new version for the first time.

Training data for the new version
1.6B
More than 1.6 billion tokens of pre-training text and training examples combined.
143KPre-training texts
2.7MTraining examples
21KTool-use examples
13,500Held-out evaluation questions
As of October 2026

Our raw archive holds more than 141,000 files from 82 source families, with a total size of about 82 GB. For MINA's pre-training we prepared 143,000 clinical texts totalling about 755 million tokens. The training set holds 2.7 million examples totalling about 908 million tokens. Together, the pre-training text and the training examples contain more than 1.6 billion tokens.

Scope of the training data

143K
Pre-training texts
The clinical text prepared for pre-training totals about 755 million tokens.
21
Clinical specialties
The training examples cover 21 clinical specialties.
21K
Tool-use examples
These examples were prepared to teach MINA to use drug, SUT and coding tools correctly.
13,500
Evaluation questions
The questions in 98 evaluation sets were kept separate from the training data.

The 21 clinical specialties describe the scope of the training data prepared for the new version. We explain on the MINA page how today's MINA v3.7 was adapted to the language of 14 clinical specialties.

Where the data comes from

The data comes from Ministry of Health guidelines, Turkish peer-reviewed medical literature, summaries of product characteristics for medicines, SUT and SGK regulations, national health coding systems, medical device registries, open-access international clinical guidelines and publications of international health authorities.

In October 2026, more than 262,000 drug labels and medical device records from the US Food and Drug Administration (FDA) were also added to the archive.

How we prepared the data

Collecting the data is only the beginning of the work. We cleaned every document, removed duplicates and recorded which source each training example came from. We checked that none of the evaluation questions appears in the training data. This work takes time, experience and a strong team.

We are putting all this effort into an AI model that will work alongside Turkish physicians with knowledge of Turkish clinical language and Türkiye's regulations.

You can read how MINA was built on the MINA page. If you would like to see it in your own institution, you can request a POC.

The figures refer to the training data prepared as of October 2026. Archive figures come from the count of 25 October 2026. The FDA records were added to the archive and are not included in the training example and token figures. Token counts are estimates. The example and token counts of the training set include examples used more than once to increase their weight in training.

SSS

Which MINA version do these figures refer to?

The figures refer to the training data prepared as of October 2026 for MINA's new version. The evaluation results published on this site refer to MINA v3.7.

How were the evaluation questions kept separate from the training data?

The 13,500 questions in 98 evaluation sets were not added to the training data. After the training data was prepared, we also checked that none of the evaluation questions appears among the training examples.

What is the difference between the 21 clinical specialties and the 14 specialties on the MINA page?

The 21 clinical specialties describe the scope of the training data prepared for the new version. The 14 specialties on the MINA page are the specialties whose language today's MINA v3.7 was adapted to.