Daring-Anteater is a comprehensive dataset for instruction fine-tuning released by NVIDIA in 2024, with the related research paper titled "HelpSteer2: Open-source dataset for training top-performing reward models", aimed at promoting reproducibility and training high-performance reward models.

The dataset contains approximately 95,000 samples, most of which were synthetically generated by NVIDIA's proprietary models and Mixtral-8x7B-Instruct-v0.1, with the remaining samples sourced from the commercially friendly subsets of FinQA, wikitablequestions, and Open-Platypus. The dataset covers a wide range of tasks and scenarios, primarily used for supervised fine-tuning (SFT) and subsequent preference alignment. The dataset is open-sourced under the CC-BY-4.0 license, supporting commercial use.

Dataset Composition

The dataset includes four main fields:

• conversations: Dialogue turns between the user and the assistant
• mask: Turns where loss is not computed (default is "User")
• system: System prompt
• dataset: Data source

The dataset consists of the following subsets:

• synthetic_conv: 82,450 samples
• synthetic_roleplay: 2,996 samples
• synthetic_math: 3,000 samples
• synthetic_precise_instruction_following: 1,500 samples
• synthetic_json_format_following: 1,499 samples
• synthetic_complex_instruction: 1,500 samples
• open_platypus_commercial: 6,000 samples
• FinQA: 300 samples
• wikitablequestions: 287 samples

Citation

@misc{wang2024helpsteer2,
      title={HelpSteer2: Open-source dataset for training top-performing reward models}, 
      author={Zhilin Wang and Yi Dong and Olivier Delalleau and Jiaqi Zeng and Gerald Shen and Daniel Egert and Jimmy J. Zhang and Makesh Narsimhan Sreedhar and Oleksii Kuchaiev},
      year={2024},
      eprint={2406.08673},
      archivePrefix={arXiv},
      primaryClass={id='cs.CL' full_name='Computation and Language' is_active=True alt_name='cmp-lg' in_archive='cs' is_general=False description='Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.'}
}