Hatim2221/Fikr-7B-Reasoning

Fikr-7B-Reasoning: Structured Arabic Reasoning

Fikr-7B-Reasoning is a 7.6 billion parameter model, fine-tuned from Qwen/Qwen2.5-7B-Instruct, with a 32768 token context length. Its primary innovation lies in enforcing structured, multi-step Chain-of-Thought (CoT) reasoning using native <think>...</think> tokens, specifically for Arabic language tasks.

Key Capabilities & Features

• Structured Reasoning: Strictly separates intermediate logical steps within <think>...</think> tags from the final parsed response, enhancing clarity and interpretability of the reasoning process.
• Arabic Mathematical Fluency: Demonstrates improved formatting stability for multi-step arithmetic, fraction reductions, and currency conversions in native Arabic phrasing.
• Targeted Fine-tuning: Utilizes Supervised Fine-Tuning (SFT) to guide the model towards generating explicit, step-by-step thought processes.

Performance & Considerations

Evaluated on the Arabic-GSM8K benchmark, Fikr-7B-Reasoning achieved 71.95% accuracy, a minor regression of -0.45% compared to its base model, Qwen/Qwen2.5-7B-Instruct (72.40%). This trade-off is common in initial SFT reasoning models, where enforcing structured thought can sometimes slightly impact raw exact-match accuracy, often requiring further Reinforcement Learning (RL) for alignment. The model shows strong performance in direct calculations and single-step problems, though it may struggle with complex logical "trap" questions without further alignment.

Recommended Use Cases

• Arabic Mathematical Problem Solving: Ideal for applications requiring detailed, step-by-step solutions to arithmetic, algebra, and financial calculations in Arabic.
• Educational Tools: Can be used to generate explanations for mathematical and logical problems, helping users understand the reasoning process.
• Structured Output Generation: Suitable for scenarios where a clear, auditable trace of the model's thought process is required before a final answer.