MARS-v0.2: Enhanced Reasoning and Turkish Language Model

MARS-v0.2 is the second iteration of models from Curiosity Technology, leveraging the robust Meta Llama 3.1 8B Instruct as its base. This 8 billion parameter model has been fine-tuned to significantly improve its reasoning and problem-solving capabilities, particularly in mathematical contexts.

Key Enhancements & Capabilities

• Mathematical Reasoning: The model's training incorporated a more comprehensive dataset with an increased focus on mathematical data, aiming to boost its analytical and problem-solving skills.
• Turkish Language Processing: Curiosity Technology continues its dedication to Turkish language support, utilizing a blend of proprietary in-house Turkish datasets and a broad selection of translated open-source datasets for fine-tuning.
• Foundation: Built on the high-performance Llama 3.1 8B Instruct architecture, providing a strong base for its specialized capabilities.
• Training: The model underwent 3 days of training on 4xA100 GPUs using LoRA fine-tuning methods.

Ideal Use Cases

• Mathematical Problem Solving: Suited for applications requiring advanced numerical and logical reasoning.
• Turkish Language Applications: Excellent for tasks involving generation, understanding, or translation within the Turkish language.
• Instruction Following: Benefits from its Llama 3.1 Instruct base, making it effective for conversational and instruction-based tasks.