Grounded Turkish Data: not just more data, but more traceable data

Why did DEVİM need a new data track?

CONTENT experiments exposed two problems: the model could exploit surface shortcuts, and the evaluator could sometimes fail to separate real capability from copying or response-format behavior.

The response was not simply "generate more Turkish tokens." The new data track was designed to make it possible to trace which answer comes from which evidence, keep training and evaluation material separated, and measure behavior when the evidence does not support an answer.

Source frame

The verified source frame has the following aggregate scale:

Sequestered works do not enter training. The purpose is not only to reduce contamination risk, but to preserve material that can remain genuinely unseen for later evaluation.

What does grounding mean here?

A supervision example is not treated as an isolated question-answer pair. The evidence supporting the answer, source provenance and validation state are tracked together.

That allows the project to ask more than "did the model give the expected answer?" It can also ask:

• is the answer supported by the supplied evidence?
• is the evidence training-side material?
• is evaluation material sequestered?
• when the answer is absent, can the model abstain?

Answer-present and answer-absent

The system is not designed only around examples that contain an answer. Answer-absent cases, where the evidence does not support the requested answer, are deliberately represented as a separate condition.

This matters for hallucination and epistemic discipline. The target is not a model that says something for every prompt, but one that can learn insufficient evidence behavior when appropriate.

Why provenance matters

The grounded corpus is not only training material; it is also a research instrument. If a metric rises unexpectedly, the project should be able to trace which source, passage, generator and validation path contributed to that behavior.

Hashes, manifests, shard accounting and source registries are therefore not merely implementation details. They form part of the scientific provenance of the experiment.

What remains private?

Under the controlled-open policy, aggregate methodology and results can be public, while rights-restricted source text, sequestered evaluation items, complete corpus contents, private generation prompts, hidden holdouts and strategically sensitive exact recipes remain private.

The goal is not to hide evidence. It is to separate the evidence needed to substantiate a public claim from the private research core.

Why this matters scientifically

The value of this data track is not only corpus size; it is diagnostic power. If fresh relation transfer still fails under grounded, counterbalanced supervision, explanations such as "the data was simply dirty" or "the model never learned the answer register" become easier to rule out.