berkbirkan/gemma-3-1b-turkish-seo-reasoning-lora

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Jul 28, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The berkbirkan/gemma-3-1b-turkish-seo-reasoning-lora is a 1 billion parameter Gemma 3 1B IT base model fine-tuned by berkbirkan using LoRA for Turkish SEO auditing tasks. It is specifically designed to analyze page or site evidence, generate a brief justification, an SEO decision, and actionable recommendations. This model excels at evidence-based SEO question answering and reasoning in Turkish, demonstrating significant performance improvements over its base model on a custom benchmark.

Loading preview...

Model Overview

This model, berkbirkan/gemma-3-1b-turkish-seo-reasoning-lora, is a LoRA fine-tuned version of the Gemma 3 1B IT base model, adapted for Turkish SEO auditing tasks. It processes given page or site evidence to produce a concise justification, an SEO decision, and actionable recommendations. The fine-tuning was performed using Unsloth and TRL on a custom berkbirkan/turkish-seo-reasoning dataset, comprising 1,080 training examples.

Key Capabilities

  • Evidence-based SEO Reasoning: Generates justifications, decisions, and recommendations based solely on provided evidence.
  • Turkish Language Support: Specifically trained and optimized for Turkish SEO contexts.
  • Structured Output: Designed to produce answers in a consistent format: "Gerekçe: ...\n\nKarar: ...\n\nÖneriler: - ... - ..."
  • Improved Performance: Achieves a +10.2781 absolute score and 85.97% relative improvement over the Gemma 3 1B IT Base model on a 120-question benchmark covering 12 SEO categories.

Use Cases

  • SEO Auditing: Automating initial assessments of web page or site SEO health.
  • Content Optimization: Providing recommendations for improving content based on SEO principles.
  • Educational Tool: Demonstrating how to evaluate SEO factors and formulate responses based on Google Search Central guidelines.

Limitations

  • The training was a pilot fine-tune of only 30 steps, not a full epoch.
  • No validation split or validation loss was used during training.
  • The automated benchmark relies on lexical similarity, which may not fully capture semantic correctness.
  • The model does not crawl live sites or access Search Console data, and its recommendations do not replace expert review.

Finetunes
1 models