cumhuronat/gemma-4-12b-cetvel-merged

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 19, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The cumhuronat/gemma-4-12b-cetvel-merged model is a 12 billion parameter Gemma-4 variant fine-tuned by Cumhur Onat for Turkish language tasks. It is specifically optimized for a range of Turkish natural language processing tasks, including question answering, multi-choice question solving, classification, natural language inference, translation, summarization, and grammatical correction. This model demonstrates significant performance improvements on the CETVEL benchmark, achieving a +18.38 overall score increase compared to its base model.

Loading preview...

Model Overview

The cumhuronat/gemma-4-12b-cetvel-merged is a 12 billion parameter model derived from google/gemma-4-12B, fine-tuned by Cumhur Onat. This merged model is the result of a 3-epoch LoRA training specifically targeting Turkish CETVEL tasks, designed for prompt/completion formats rather than chat or instruction-tuned interactions.

Key Capabilities and Performance

This model is engineered to excel in various Turkish NLP tasks, showing substantial improvements over the base Gemma-4 model. It was trained on the cumhuronat/cetvel-training-data dataset, which includes CETVEL-formatted training splits, Turkish/English instruction data, synthetic task examples, and deterministic answer-format examples.

CETVEL Benchmark Highlights:

  • Overall Score: Achieved 51.22, a +18.38 point increase (55.98% relative improvement) compared to the baseline's 32.84.
  • Grammatical Correction (GEC): Saw the most significant gain, improving from 40.22 to 91.34 (+51.12).
  • Natural Language Inference (NLI): Improved from 32.64 to 57.43 (+24.79).
  • Text Classification (TC): Increased from 40.10 to 61.14 (+21.04).
  • Question Answering (QA): Improved from 22.57 to 42.20 (+19.62).

Intended Use Cases

This model is specifically designed for:

  • Turkish question answering
  • Multi-choice question solving
  • Text classification
  • Natural language inference
  • Translation
  • Summarization
  • Grammatical correction experiments

It is recommended for use with a bf16-supporting GPU. The model's license follows Google Gemma terms, while the pipeline code is Apache-2.0 licensed.

Finetunes
1 models