RLMIA/GRPO-3B-SEARCH-R1

RLMIA/GRPO-3B-SEARCH-R1: Qwen2.5 Instruction-Tuned Model

RLMIA/GRPO-3B-SEARCH-R1 is an instruction-tuned variant of the Qwen2.5 series, a family of large language models developed by Qwen. This specific model has 3.09 billion parameters and supports a substantial context length of 32,768 tokens, with a generation capacity of up to 8,192 tokens.

Key Capabilities and Improvements

This model builds upon the Qwen2 architecture, offering significant enhancements:

• Expanded Knowledge & Specialized Skills: Features substantially more knowledge and improved performance in coding and mathematics, leveraging specialized expert models.
• Instruction Following: Demonstrates significant improvements in adhering to instructions and generating structured outputs, particularly JSON.
• Long Text Generation: Excels at generating extended texts, capable of producing outputs over 8,000 tokens.
• Structured Data Understanding: Enhanced ability to comprehend and process structured data, such as tables.
• Multilingual Support: Provides robust support for over 29 languages, including major global languages like Chinese, English, French, Spanish, German, and Japanese.
• Resilience: More resilient to diverse system prompts, which benefits role-play and chatbot condition-setting.

Architecture and Training

The model utilizes a transformer architecture incorporating RoPE, SwiGLU, RMSNorm, Attention QKV bias, and tied word embeddings. It underwent both pretraining and post-training stages to achieve its instruction-tuned capabilities.

When to Use This Model

This model is particularly well-suited for applications requiring strong instruction following, code generation, mathematical problem-solving, and the ability to process and generate long, structured text or multilingual content. Its resilience to varied system prompts also makes it a good choice for complex conversational AI and role-playing scenarios.