ModernBERT-TR
A long-context (8,192 tokens) Turkish encoder with a family of specialized downstream models.
Projects, Models & Datasets
Models, datasets, demos and systems I've built or co-built, mostly open source.
Models
ChainFlow
A flow-matching drafter for lossless speculative decoding in vLLM: up to 2.5x faster on Qwen3.5 (4B to 27B), with Turkish-adapted checkpoints.
Cosmos-LLaVA
Turkish's first large-scale multimodal datasets along with a vision-language instruction-following model.
Turkish language models
A range of Turkish language models built on Llama, Gemma, E5, BERT, GPT and other architectures. More in the COSMOS Hugging Face organization.
Datasets
ABSA-TR
Turkish Aspect-based sentiment analysis dataset. Built from native sources in the e-commerce, supplements, and movie domains.
Guardrail-TR
About 405K Turkish prompts for safe/unsafe classification. Multi-label hazard categories covering 11 labels.
Large clean Turkish web corpus
A cleaned 125 GB Turkish pretraining corpus built from CulturaX (ICMI 2025).
Turkish-LLaVA-Pretrain
Turkish multimodal pretraining dataset behind Cosmos-LLaVA.
Turkish-LLaVA-Finetune
Turkish visual instruction fine-tuning dataset. One of the datasets behind Cosmos-LLaVA.
turkce-kitap
Turkish conversations about book covers, for vision-language training.
Projects
Optimizing LLMs for Turkish
Better data curation and training methodologies for Turkish language models.
Turkish legal decision-support system
1st place, TEKNOFEST 2025 Turkish NLP Competition.
Spatial reasoning via maze navigation
Teaching compact LLMs spatial reasoning through maze navigation (ICETAI 2026).