unicorn-team/Unicorn-VL-R3

Overview

Unicorn-VL-R3 is an 8 billion parameter vision-language model developed by unicorn-team, fine-tuned from the Qwen3-VL-8B base model. It is specifically designed and optimized for Vietnamese educational tasks, demonstrating strong performance on academic benchmarks.

Key Capabilities

• Specialized Training Data: Utilizes a unique synthetic data generation process, creating approximately 4,800 samples of question-response pairs with simulated thinking chains (synthetic CoT) based on 12 educational tasks from Gemini APIs. This approach aims to optimize the model's ability to solve school-related problems with logical reasoning.
• Vietnamese Language Focus: Incorporates about 5,000 high-quality Vietnamese samples filtered from the Dolci-SFT dataset, ensuring strong performance in the Vietnamese language context.
• Vision-Language Integration: As a VL model, it is capable of processing both visual and textual inputs, making it suitable for multi-modal educational content.
• Performance on VMLU: Achieves a VMLU score of 74.87, outperforming its base model Qwen3-VL-8B (74.10) and other Unicorn-R3 variants.
• Optimized for Reasoning: The synthetic CoT generation strategy is designed to improve the model's reasoning and problem-solving skills, particularly for academic questions.

Good For

• Applications requiring a robust Vietnamese vision-language model for educational content.
• Tasks involving academic problem-solving and reasoning in Vietnamese.
• Use cases where strong performance on the VMLU benchmark is critical.