Dataset Overview

ChatQA 2 Long SFT Data is a dataset released by NVIDIA in 2024 that focuses on long-context conversational question answering and Retrieval-Augmented Generation (RAG) capabilities. Related research outcomes can be found at 「ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities」. It aims to enhance the performance of large language models when handling tasks involving long contexts and complex multi-turn conversations.

The long_sft subset is derived from samples generated using GPT-4 within LongAlpaca12k, Open Orca, and Long Data Collections; while the NarrativeQA_131072 subset was synthetically generated based on NarrativeQA by adding paragraphs related to given ground-truth summaries. This dataset serves as the full long-context supervised fine-tuning (SFT) training data for the ChatQA-2 model and primarily consists of two subsets: long_sft and NarrativeQA_131072. The data format is JSON, containing both training and test sets used during the instruction tuning phase of the model.

Dataset Composition

The dataset includes the following main configurations (subsets):

• long_sft: Long-context supervised fine-tuning data sourced from existing datasets, designed to improve the model's understanding and generation abilities in long-text dialogues.

  • Training set: long_sft/long_sft_QA_train.json
  • Test set: long_sft/long_sft_QA_dev.json

• NarrativeQA_131072: Long-context QA data synthesized based on the NarrativeQA dataset, expanding contextual information by incorporating relevant paragraphs.

  • Training set: NarrativeQA_131072/NarrativeQA_131072_QA_train.json
  • Test set: NarrativeQA_131072/NarrativeQA_131072_QA_dev.json

Citation

@article{xu2024chatqa,
  title={ChatQA 2: Bridging the Gap to Proprietary LLMs in Long Context and RAG Capabilities},
  author={Xu, Peng and Ping, Wei and Wu, Xianchao and Liu, Zihan and Shoeybi, Mohammad and Catanzaro, Bryan},
  journal={arXiv preprint arXiv:2407.14482},
  year={2024}
}