Avelina/UltraSteer-v0

UltraSteer是一个大规模的对话数据集，包含单轮和多轮对话，并带有细粒度的标签。这些标签由Nvidia的Llama2-13B-SteerLM-RM奖励模型生成，使用了NeMo Aligner框架。数据集中的每个助手回复都根据多个属性进行评分，包括质量、毒性、幽默感、创造力、帮助性、正确性、连贯性、复杂性和详细程度。数据集还进行了因果去重处理，确保每个对话的最后一个助手回复都有标签且唯一。数据来源于多个高质量的对齐数据集和聊天数据集，并进行了过滤和去重处理。

UltraSteer is a large-scale dataset of single- and multi-turn dialogue with fine-grained labels produced by Nvidias Llama2-13B-SteerLM-RM reward model using the NeMo Aligner framework. Each assistant turn is rated with attributes such as quality, toxicity, humor, creativity, helpfulness, correctness, coherence, complexity, and verbosity. The dataset has undergone causal deduplication to ensure that the final assistant message in every conversation is labeled and unique. The data is sourced from multiple high-quality alignment datasets and chat datasets, and has been filtered and deduplicated.