Turkish models from ufak AI
ufakzeka
We build small, open Turkish models, trained from scratch or on our own base model, and a benchmark for decision models.
Models43 models, 1 benchmark
Models and benchmark
ufakzeka-1
Chat model
A 151M parameter chat model trained from scratch for Turkish.
- 151Mparameters
- $286total spend
ufakzeka-karar
Decision model
It reads Turkish text and answers choice, score and yes/no questions with a probability for every option and one not-sure signal per question.
- 182,494,466parameters, embedding matrix included
- 102.5 msmedian per question, one thread on an Apple M1 CPU
HakemBench
Benchmark
A fully open benchmark that measures typed decisions in Turkish across seven tracks.
- 4,275questions
- 2,346items
ufakzeka-kulak
Speech recognition model
It writes Turkish speech out as text, with capitals and punctuation, offline on a CPU.
- 31,564,935parameters
- 0.0921word error rate, FLEURS Turkish test
In commontwo models, one base
One base model
ufakzeka-1 and ufakzeka-karar come from ufakzeka-1-base, a base model trained from scratch for Turkish. ufakzeka-1 is its fine-tuned chat version, with 151M parameters; ufakzeka-karar starts from the same base weights and adds a decision head. HakemBench measures ufakzeka-karar and other models on Turkish decision questions with the same harness. The 151M of ufakzeka-1 leaves out the embedding matrix; ufakzeka-karar’s 182,494,466 parameters include it (151,037,186 without it). ufakzeka-kulak is not built on that base; it is a speech recognition model of 31,564,935 parameters, trained separately from scratch on speech recordings.
- 3open models
- 1benchmark
- ufakzeka-1-basebase of the chat and decision models
- Apache-2.0model licence