ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities

TL;DR

ChatQA 2 introduces an extended-context Llama 3.0-based model with 128K tokens, significantly improving long context understanding and retrieval-augmented generation, outperforming existing models on ultra-long and RAG tasks.

Contribution

The paper presents a novel training recipe and instruction tuning process to extend Llama 3.0's context window to 128K tokens and enhance RAG capabilities, outperforming state-of-the-art models.

Findings

Outperforms GPT-4-Turbo and other models on ultra-long tasks.

RAG with larger top-k chunks improves performance over direct long-context methods.

Open-sourced model, training data, and evaluation setup.

Abstract

In this work, we introduce ChatQA 2, an Llama 3.0-based model with a 128K context window, designed to bridge the gap between open-source LLMs and leading proprietary models (e.g., GPT-4-Turbo-2024-04-09) in long context understanding and retrieval-augmented generation (RAG) capabilities. These two capabilities are complementary to each other and essential for LLMs to process large volumes of information that cannot fit into a single prompt. We present a detailed continued training recipe to extend the context window of Llama3-70B-base from 8K to 128K tokens, along with a three-stage instruction tuning process to enhance the model's instruction-following, RAG performance, and long-context understanding capabilities. Our results demonstrate that the Llama3-ChatQA-2-70B model outperforms most existing state-of-the-art models, including GPT-4-Turbo-2024-04-09, Qwen2-72B-Instruct, and Llama3.1-70B-Instruct, on ultra-long tasks beyond 100K tokens, as well as on the RAG benchmark using only a 4K context window, showing the strong long context capability across varying sequence lengths. We further provide extensive comparisons between direct long-context and RAG solutions using the same state-of-the-art long-context LLMs. Interestingly, we find that the performance of strong long-context LLMs using RAG improves when retrieving a larger number of chunks. With a large set of top-k chunks, RAG consistently outperforms direct long-context solution using the same state-of-the-art long-context models (e.g., Llama3-ChatQA-2-70B and Qwen2-72B-Instruct) on both 32K and 128K benchmarks. We open-source the model weights, training data, and the evaluation setup for the for the community: https://chatqa2-project.github.io/