Hi, I’m Ibrahim Sait. I’m a 21-year-old Computer Science student originally from Türkiye. After moving to Warsaw at 18, I have now settled in Luxembourg, where I’m continuing my Computer Science education at the University of Luxembourg.
Thanks to the incredibly ambitious and forward-thinking friends I met in high school, I entered the world of business and startups early through Mobil Çelebi. Since then, neither the way I see life nor the goals I set for myself have ever been the same.
My philosophy is rooted in sacrifice and freedom. I deeply value spending time with people who are willing to give up their own comfort for something greater, and I hope to live a life worthy of them.
No tested change beat the RTX 3090 baseline. The best—and only accepted—result remained 1.245730 prediction loss, so the practical outcome is to keep the baseline configuration unchanged.
The best measured prediction loss was 1.113377, down from the 1.248332 baseline—an improvement of about 10.8%. However, this result came from reducing the training batch size, not from a sampling-order or curriculum…
The best measured configuration reduced validation prediction loss from 1.247465 to 1.206223—an absolute improvement of 0.041242, or about 3.3%. The decisive result came from using a seven-layer model, which performed…
The best configuration reduced validation prediction loss from 1.247607 to 1.208420—an absolute improvement of 0.039187, or about 3.14%. The effective recipe was to tie the input and output token weights, lower the…
No rotary-position change improved the benchmark. The best validation result remained the RTX 3090 baseline of 1.248862 bits per byte; the only alternative that completed evaluation was worse, while the layer-wise…
Beyefendi — Qwen3.5-2B tabanı üzerinde küçük bir Türkçe Wikipedia continued pretraining aşaması ve temizlenmiş Türkçe Aya instruction verisiyle SFT uygulanmış, yaklaşık 2B sınıfında Türkçe öncelikli bir sohbet modelidir.
Beyefendi-GGUF — Beyefendi modelinin llama.cpp ve Ollama ile yerel çalıştırmaya yönelik F16 ve Q4_K_M biçimlerini içeren nicemlenmiş dağıtım deposudur.
Beyefendi-v2 — Qwen3.5-9B için Türkçe ve İngilizce verilerle eğitilmiş, tamamlanmış ve hashleri doğrulanmış bir public-development LoRA checkpoint’idir.
Beyefendi-v2-GGUF — Beyefendi v2 geliştirme checkpoint’inin GGUF biçimindeki yerel çalışma artefaktlarını yayımlamak için açılmış public model deposudur.