Vi-Qwen2-7B-RAG: Specialized for Vietnamese RAG Tasks

Vi-Qwen2-7B-RAG is a 7.6 billion parameter large language model developed by AITeamVN, fine-tuned from the Qwen2-7B-Instruct base model. Its primary focus is to enhance performance for Retrieval Augmented Generation (RAG) tasks specifically in the Vietnamese language, leveraging a substantial Vietnamese dataset for training. The model supports an impressive context length of up to 131072 tokens, with RAG-specific contexts up to 8192 tokens.

Key Capabilities

• Noise Robustness: Effectively extracts useful information from noisy documents.
• Negative Rejection: Accurately declines to answer questions when necessary knowledge is not present in the retrieved documents.
• Information Integration: Capable of answering complex questions that require synthesizing information from multiple documents.
• Context Identification: Achieves approximately 99% accuracy in determining if a context contains the answer to a question.
• General Chatbot: Can also function as a regular chatbot, supporting continuous conversation with input context.

Performance & Benchmarks

The model's RAG capabilities were evaluated using a custom, human-scored dataset, EvalRAGData, demonstrating strong performance. Additionally, it achieved an average score of 56.04 on the VMLU leaderboard, showcasing its general Vietnamese language understanding.

Limitations

Due to its specialized design for RAG tasks, the model may exhibit limitations in areas such as political or social questions, and might occasionally display biases or inappropriate views.

Good for

• Developers building RAG systems requiring high accuracy in Vietnamese.
• Applications needing robust information extraction from diverse documents.
• Use cases where the model must discern relevant information and refuse to hallucinate.