Aknar–Mutlu Model Family

A design specification for training a Turkish pathology foundation model by LoRA-adapting OpenMidnight on a local multi-organ cohort — the first project here that sets out to build an encoder rather than consume one.

Purpose

Every other imaging project in this wiki treats a foundation model as something you download. This one proposes to make one. Aknar v1 is specified as a continued-pretrained variant of OpenMidnight (Apache-2.0, ViT-G/14, 1.1B parameters) adapted to a Turkish multi-organ pathology cohort by LoRA fine-tuning — rank ≈ 16, roughly 10–30M trainable parameters, backbone frozen. The stated success criterion is to beat OpenMidnight base on Turkish tasks, and the release intent is institutional and hospital-internal.

The naming is not decoration and the repository treats it seriously enough to give it two appendices. The model line honours Hamdi Suat Aknar (1873–1936), founder of modern Turkish pathology and the first professor of pathology in Turkey; the family also honours Kâmile Şevki Mutlu (1906–1987), Aknar’s assistant and Turkey’s first female pathology specialist.

It is explicitly a design specification and not an implementation plan — the document says so in its own scope note, and records that implementation will follow as a separate deliverable once the design is approved. Four commits, last 2026-06-02.

Data used

A Turkish multi-organ pathology cohort, ~1,000–5,000 WSIs for Phase 1. Nothing in the tracked material specifies which organs, which scanners, or where the slides come from, which is the same acquisition-metadata gap this wiki records against every other imaging project — and it matters more than usual here, because the whole premise is that a local corpus teaches the encoder something the base model does not know. If the corpus turns out to be one scanner’s output, the thing learned may be the scanner. [unverified]

Methods

The design is shaped by one hard constraint, and states it: an A4000 with 16 GB is the only compute available. A 1.1B-parameter ViT-G/14 cannot be fully fine-tuned in that budget, so the backbone is frozen and adaptation happens through low-rank adapters — which is why LoRA is the method rather than a preference. That is the honest way round: the constraint chose the technique, and the document says so instead of presenting LoRA as independently optimal.

Two things about this are worth drawing out.

The base model choice is well made. OpenMidnight is on the Foundation Model Cache drive already, so Phase 1 needs no new download. It is also Apache-2.0, which for a model the group intends to adapt and release internally is a materially easier starting point than the alternatives — most of the strongest encoders in that cache carry terms that would complicate distributing an adapted version.

The success criterion is a real one. “Beat OpenMidnight base on Turkish tasks” names a baseline and a population, which is more than most of the model work recorded here does. It is also the correct comparison: the question is not whether Aknar v1 is good, but whether local adaptation bought anything over the model it started from. The obvious gap is that the tasks are not enumerated, and until they are, “Turkish tasks” cannot be evaluated. Choosing two or three concrete downstream tasks with existing local labels — the kidney cohort behind PinkKidney is the best-controlled one available — would make the criterion testable.

Current state / open questions

Design approved and written up on 2026-05-30; nothing since 2026-06-02. Private repository.

Only the rendered book is in version control. All 35 tracked files sit under _book/ — the Quarto HTML output. The .qmd sources that produced them are not committed. So the repository publishes the design and does not preserve the means of regenerating or amending it, which is the inverse of what a version-controlled design document is for. Same shape as PinkKidneyWeb, though there it is deliberate and here it looks accidental.

• Which downstream tasks define success? Unspecified, and the criterion is unusable without them. This is the cheapest thing to fix and it gates the whole evaluation.
• Is continued pretraining on 1–5k WSIs enough to move a 1.1B-parameter encoder? That is the real scientific risk. Published continued-pretraining results generally use corpora one to two orders of magnitude larger; LoRA reduces the trainable parameters but not the amount of signal in the data. [unverified] — and worth searching for a precedent at this scale before spending the compute.
• What would count as “Turkish” about the result? If the gain comes from organ mix, staining protocol or scanner rather than anything population-specific, the model is a local-domain adaptation rather than a Turkish one. Both are useful; they are different claims, and only one of them justifies the name.
• Phase 2 onward is referenced as a multi-stage roadmap but not specified here. [unverified]

Related: Foundation Models in Pathology — this is the concept approached from the training side rather than the consuming side, which nothing else here does. Related: Foundation Model Cache — OpenMidnight, the base model, is already on the drive. Related: PinkKidney — the best-controlled local cohort, and the natural place to define a concrete downstream task for the success criterion. Related: Scanner and Stain Variability — the confounder that decides whether a locally adapted encoder learned the population or the equipment. Related: Whole Slide Imaging — the 1–5k WSI corpus is the input, and its acquisition metadata is unrecorded.

Derived from: repository read 2026-07-29 — the rendered book at _book/, specifically docs/aknar_model_workflow.html (Phase 1 design, read for scope, executive summary and constraints) and docs/about_aknar.html; git ls-files, git log and gh repo view for what is tracked, currency and visibility.