devim-model · research
Why Turkish-first?
A short framing of why Devim treats Turkish as a starting condition of model research rather than a localization layer added later.
Language is not only an output format
In Devim, Turkish is not treated as a localization layer added to an English model. Data selection, tokenizer behavior, training distribution, evaluator design and generation form are studied together.
What are we trying to separate?
Poor sentence formation and poor relation binding are not the same failure. We therefore track at least two axes:
- FORM: sentence construction, repetition, termination behavior and response form.
- CONTENT: relation binding, instruction following, generalization and abstention when appropriate.
This separation helps diagnose the source of failure before scaling the model.
Research principle
Scale is not an explanation. Before increasing parameter count, we try to isolate data, evaluator and training-schedule failures through controlled tests.
- Version
- v1
- Evidence
- 0
- Source
- Git · en/research/why-turkish-first.md
DEVİM publications preserve revision history and distinguish public evidence from internal work.