Allenda/Qwen2.5-7B-Instruct-cognify-BreaK

Qwen/Qwen2.5-7B-Instruct fine-tuned for CognifyChallenge 2026 (team BreaK): given a synthetic patient record, a dementia classifier's verdict and its top-5 SHAP attributions, write a short clinical note explaining the model's reasoning, plus a structured list of the drivers the note relies on.

What it was trained on

650 examples: 450 cases from the practice split of the competition's own cohort (cohort_dev.json), plus 200 sparsified copies of those same cases.

The targets are not human-written. They are generated by a deterministic program that assembles the note the competition's scoring function implies — verdict and confidence band in the question bank's own wording, the five attributions in rank order with the direction each one gives, values quoted only where the record can support them, and the findings the record reports as present so that what it reports as absent follows by elimination. So this is distillation of a program into a model, not a model taught medicine.

The sparsified copies exist because the two splits differ in the tail: the practice split's records carry at least 21 citable fields and the held-out split reaches down to 7. Trained on the practice range alone, the model wrote a correct note on 59 of 60 manufactured sparse cases and fell into a degenerate repetition loop on the sixtieth. With the sparse copies in the training set it is 60 of 60.

The held-out slice of 150 cases was excluded from training, and the sparsified copies were built only from training cases. The competition's test split was never read.

Intended use

This model is for one task, in one prompt format, on synthetic data. It does not diagnose, and it should not be pointed at real patient records. Every patient in its training data is synthetic.

Prompt format

It expects the rendering the competition harness produces — the patient record, the model output, the top 5 attributions and the citable-fields JSON, followed by a short instruction — and replies with a single JSON object holding note and drivers. Outside that format it will still answer, but nothing about its behaviour there was measured.