技术规格

GreenMind-14B-R1

GreenMind-14B-R1 is a 14.7 billion parameter Vietnamese reasoning model developed by GreenNode. It is a dense, decoder-only transformer derived from the Qwen2.5-14B-Instruct base architecture. The model is specifically engineered for multi-step logical reasoning and high-fidelity text generation in the Vietnamese language, addressing common limitations such as language mixing and factual drift in long-form reasoning chains. By implementing Chain-of-Thought (CoT) methodologies, GreenMind is designed to decompose complex queries into intermediate logical steps before producing a final response.

The model utilizes a specialized fine-tuning strategy known as Group Relative Policy Optimization (GRPO), which optimizes the reasoning process while maintaining computational efficiency. This training approach is augmented by a curated Vietnamese instruction dataset consisting of over 55,000 samples spanning cultural, legal, and educational domains. To ensure linguistic consistency, the training pipeline incorporates specific reward functions and Sentence Transformer-based verification to prevent the intrusion of non-Vietnamese characters and to preserve the factual integrity of the reasoning trajectories.

Optimized for deployment via NVIDIA NIM, GreenMind-14B-R1 is intended for enterprise-grade applications including legal and financial assistants, context-aware conversational agents, and complex document retrieval systems. The architecture supports a context length of up to 131,072 tokens for input processing, with a maximum generation limit of 8,192 tokens. Its integration of modern transformer techniques like RoPE position embeddings and SwiGLU activation makes it a technically sophisticated tool for localized AI infrastructure in Vietnam.

没有可用的 GreenMind-14B-R1 评估基准。