ameer4wisam/gemma-iraqi-10k-merged

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 29, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

The ameer4wisam/gemma-iraqi-10k-merged model is a 12 billion parameter Gemma 4 instruction-tuned model, developed by ameer4wisam, specifically fine-tuned for the Iraqi dialect. This merged model integrates a LoRA adapter trained on 14,000 Iraqi examples and general instructions, preserving both vision and audio towers by using AutoModelForImageTextToText. It excels in generating responses in colloquial Iraqi Arabic and is suitable for applications requiring dialect-specific conversational AI.

Loading preview...

Overview

This model, ameer4wisam/gemma-iraqi-10k-merged, is a fully merged version of google/gemma-4-12B-it, incorporating a LoRA adapter fine-tuned on the Iraqi dialect. Unlike typical merges, it utilizes AutoModelForImageTextToText to ensure the preservation of the base model's vision and audio capabilities, which would otherwise be lost with text-only merging classes. The training data includes 14,000 Iraqi dialect examples, supplemented with approximately 15% general instructions from the Aya dataset.

Key Capabilities

  • Iraqi Dialect Proficiency: Specialized in understanding and generating colloquial Iraqi Arabic, particularly leaning towards central and southern Iraqi dialects.
  • Multimodal Support: Retains the ability to process both text and image inputs, a unique feature preserved through its merging methodology.
  • Agent Behavior Training: Includes specific training for agent-like behaviors, such as confirming availability, declining unavailable brands, deferring calculations, and using tool calls, based on 690 examples.

Good For

  • Conversational AI in Iraqi Arabic: Ideal for chatbots or virtual assistants designed to interact naturally in the Iraqi dialect.
  • Multimodal Applications: Suitable for use cases where both text and image understanding are required within an Iraqi linguistic context.
  • Specialized Agent Systems: Can be adapted for roles requiring specific interaction patterns, like sales assistants, where defined responses and tool invocation are necessary.

Limitations

  • The Iraqi dialect data primarily reflects central and southern variations, not a unified representation.
  • Product and pricing examples in the training data are illustrative and not based on real catalogs.
  • Performance has been evaluated internally and not against published general benchmarks.

Finetunes
1 models