AITeamVN/Vi-Qwen2-1.5B-RAG

Overview

AITeamVN's Vi-Qwen2-1.5B-RAG is a 1.5 billion parameter language model, a smaller variant of the Vi-Qwen2-7B-RAG, fine-tuned from the Qwen2-Instruct base model. Its primary focus is on enhancing performance for Retrieval Augmented Generation (RAG) tasks within the Vietnamese language. The model is trained on a Vietnamese dataset to improve its language processing capabilities and RAG efficiency.

Key Capabilities

• Noise Robustness: Effectively extracts useful information from noisy documents.
• Negative Rejection: Refuses to answer questions when necessary knowledge is not present in retrieved documents.
• Information Integration: Answers complex questions requiring information synthesis from multiple documents.
• Context Identification: Accurately determines if a context contains the answer to a question (approx. 99% accuracy).
• Chatbot Functionality: Can also be used for general chatbot interactions, including continuous conversations with input context.

Benchmarks

The model's RAG performance was evaluated using a human-scored, custom-created Vietnamese dataset, EvalRAGData. While specific scores for the 1.5B model are not detailed, the family of models demonstrates strong RAG capabilities. VMLU leaderboard benchmarks are also provided for the 7B variant, indicating general language understanding.

Limitations

• Specialized for RAG: May have limitations in accuracy for questions related to politics or social issues.
• Potential Biases: Like many LLMs, it may exhibit biases or inappropriate viewpoints.