kiki-ailab/Qwen2.5-3B-Instruct-KAI
The kiki-ailab/Qwen2.5-3B-Instruct-KAI is a 3.1 billion parameter instruction-tuned causal language model developed by kiki-ailab, based on the Qwen2.5 architecture. This model is specifically optimized for Vietnamese language understanding and generation tasks, including reading comprehension, information extraction, question answering, and summarization. It demonstrates significant performance improvements over its base model on Vietnamese benchmarks like VMLU, ViSquad, ViDrop, and ViDialog.
Loading preview...
kiki-ailab/Qwen2.5-3B-Instruct-KAI: Optimized for Vietnamese NLP
The kiki-ailab/Qwen2.5-3B-Instruct-KAI is a 3.1 billion parameter instruction-tuned model derived from the Qwen2.5 architecture. It is part of a collection of models fine-tuned by kiki-ailab with a primary focus on enhancing performance for Vietnamese language tasks.
Key Capabilities & Optimizations
- Vietnamese Language Proficiency: This model is specifically optimized for a range of Vietnamese NLP tasks, including:
- Reading comprehension
- Information extraction
- Question answering
- Summarization
- Performance Improvements: Benchmarks on the VMLU suite (vmlu.ai) show substantial gains over the base Qwen2.5-3B-Instruct model:
- VMLU: 63.5 (+10.6) compared to 52.9
- ViSquad: 94.2 (+5.9) compared to 88.3
- ViDrop: 80.9 (+8.5) compared to 72.4
- ViDialog: 68.5 (+14.1) compared to 54.4
- ArenaHard Evaluation: On the ArenaHard benchmark, the model achieves a win rate of 38.7%, outperforming the base Qwen2.5-3B-Instruct (18.6%) and several larger models like Llama-3.2-3B-Instruct (21.2%).
Good for
- Vietnamese NLP Applications: Ideal for developers building applications that require robust understanding and generation in Vietnamese.
- Resource-Efficient Deployment: As a 3.1B parameter model, it offers a balance of performance and computational efficiency for deployment in environments with limited resources.
Limitations
- May exhibit hallucinations on culturally specific content.
- Primary focus is on Vietnamese; performance in specialized technical domains might not be optimal.