ameer4wisam/gemma-iraqi-finetune-v2

Overview

This model, ameer4wisam/gemma-iraqi-finetune-v2, is a specialized conversational AI built on google/gemma-4-E4B-it. It's fine-tuned to understand and generate responses in the Iraqi dialect, focusing on sales, services, and daily life interactions. The model is designed to provide short, direct answers, emulating the communication style of an Iraqi seller.

Key Capabilities & Features

• Iraqi Dialect Specialization: Trained on 163,429 conversations in Iraqi dialect, ensuring authentic language use.
• Optimized for Sales & Services: Excels in scenarios involving buying, selling, and general service inquiries.
• High Accuracy: Achieved a validation loss of 0.215 and a token accuracy of 93.8% during training.
• Context Retention: Demonstrates strong ability to maintain context over long conversations (up to 785 tokens), remembering customer details and previous requests.
• Deterministic Responses: Offers a deterministic mode for factual queries (prices, numbers) to prevent hallucination.

Important Usage Guidelines

To ensure optimal performance and avoid 'word salad' outputs, specific inference settings are mandatory:

• attn_implementation: Must be set to "eager" for the E4B architecture.
• max_new_tokens: Recommended at 64, as training data responses are typically short.
• repetition_penalty: Must be avoided as it penalizes common dialect phrases.
• temperature: Keep below 0.3; higher values can degrade specialized output quality.

Limitations

• Numerical Accuracy: The model is weak at calculations; prices and totals should be pre-calculated and injected.
• Fact Hallucination: Without a catalog or deterministic settings, it may generate unreliable numbers.
• Generalization: Its capabilities are specialized; performance outside sales/service contexts may vary.