Cihat Yıldız PRO

Divan, a Turkish LLM arena, is now live on Spaces. It puts two models side by side on the same Turkish prompt so that differences in reasoning, fluency and instruction-following become directly comparable instead of anecdotal. The space has two modes: an open comparison where you pick both models, and a blind mode that hides model identities until after you have read both answers — removing the brand bias that makes informal LLM comparisons unreliable. Prompts are organized into 6 categories (general knowledge, reasoning, creative writing, summarization, code, translation) with ready-made Turkish test items, and every model runs through the HF Inference API.

🤗 Demo: cihatyldz/divan-turkish-llm-arena

updated a Space 8 days ago

Kervan Lojistik RAG

🐪

Türkçe Lojistik RAG · Pure LLM vs RAG+LLM karşılaştırma

published 2 Spaces 8 days ago

Kervan Lojistik RAG

🐪

Türkçe Lojistik RAG · Pure LLM vs RAG+LLM karşılaştırma

Divan — Türkçe LLM Arena

🏛

Turkish LLM Arena - live model comparison

posted an update 9 days ago

Post

105

CV Matching Engine is now live on Spaces. It takes a PDF CV and a job description and produces an explainable relevance ranking rather than an opaque score: the pipeline runs semantic NLP over the parsed document, applies hybrid scoring that combines semantic similarity with lexical signals, and surfaces the reasoning behind each ranking decision. The goal is to make CV screening auditable — recruiters and candidates can see which parts of a profile drove the match, instead of trusting a single number. Built for Turkish and English CVs.

🤗 Demo: cihatyldz/cv-matching-engine

posted an update 12 days ago

Post

Mizan, a RAG-based Turkish disinformation fact-checker, is now live on Spaces. It is built on the Directorate of Communications (İletişim Başkanlığı) DMM dataset of 2,810 officially verified disinformation claims, released openly under CC-BY-4.0. Users enter a suspicious claim and the system retrieves the most semantically similar bulletin records, combining Turkish Sentence-BERT embeddings with a FAISS index for millisecond-level search and Qwen2.5-7B for grounded assessment. The output is a three-part verdict: a similarity-scored match indicator, the matched bulletins with their official corrections, and an LLM explanation constrained to the retrieved evidence.

📊 Data source: iletisim/dezenformasyon-bultenleri
🤗 Demo: cihatyldz/mizan-fact-checker

posted an update 14 days ago

Post

120

Çarşı, a dual-model Turkish e-commerce NER demo, is now live on Spaces. It runs Named Entity Recognition side by side with a fine-tuned Turkish BERT and Qwen2.5-7B zero-shot extraction, making the "specialized NER vs. zero-shot LLM" trade-off visible in real time. The system extracts 7 entity types from Turkish product text — PRODUCT, BRAND, PRICE, COLOR, SIZE, MATERIAL and GENDER — with color-coded highlighting over the original sentence. The fine-tuned BERT returns entities in ~10-50ms on CPU, while the LLM handles open-ended extraction through the Inference API. Trained on 8,000 synthetic Turkish e-commerce sentences with BIO-tagged token classification and no manual annotation.

🧠 BERT model: cihatyldz/carsi-bert-turkish-ecommerce-ner
📊 Dataset: cihatyldz/carsi-turkish-ecommerce-ner
🤗 Demo: cihatyldz/carsi-ecommerce-ner

updated a Space about 2 months ago

LojistikAI Kargo Asistanı

🚚

Tool-calling kargo asistanı (SQLite + Jinja2)

Cihat Yıldız PRO

AI & ML interests

Recent Activity

Organizations

cihatyldz's activity

Fayton — Türkçe Araç İlanı Çözümleme

Fayton — Türkçe Araç İlanı Çözümleme

Divan — Türkçe LLM Arena

Exporting a blind comparison for reproducible evaluation

Akçe Türkçe Bankacılık AI

Akçe Türkçe Bankacılık AI

Kervan Lojistik RAG

Kervan Lojistik RAG

Divan — Türkçe LLM Arena

LojistikAI Kargo Asistanı