NVIDIA Releases ChatQA Model with Performance Comparable to GPT-4
-
NVIDIA has launched the ChatQA model, which is said to rival GPT-4 in performance, utilizing efficient training methods like two-stage instruction tuning and enhanced context retrieval.
ChatQA is a set of conversational question-answering (QA) models capable of achieving GPT-4-level accuracy. Specifically, the development team proposed a two-stage instruction tuning method that significantly improves zero-shot conversational QA results for large language models (LLMs).
To handle retrieval in conversational QA, a dense retriever was fine-tuned on multi-turn QA datasets, delivering results comparable to using state-of-the-art query rewriting models while substantially reducing deployment costs. Notably, ChatQA-70B outperforms GPT-4 on average scores across 10 conversational QA datasets (54.14 vs. 53.90) without relying on any synthetic data from OpenAI GPT models.