Nemotron-SFT-Safety-v1-prompt-only
收藏资源简介:
Nemotron-SFT-Safety-v1-prompt-only 是一个专门从安全微调数据集 nvidia/Nemotron-SFT-Safety-v1 中提取提示(prompt)部分而创建的数据集。该数据集的核心文件是 `prompts.csv`,它包含了从源数据中提取出的每条提示记录。每条记录主要包含 `prompt`(用户提示)和 `system_prompt`(系统提示)字段。此外,如果源数据行中定义了可用的工具,记录中还会包含结构化的 `tools` 字段,其中的嵌套值以 JSON 格式编码在 CSV 单元格内。数据集还提供了两个辅助文件:`summary.md` 汇总了源数据行数、提取行数、计数差异以及提取失败的提示数量;`null_or_empty_rows.md` 则列出了提示提取结果为 null 或空白的行索引。数据集规模为 45,144 条成功提取的提示记录,有 1 条提取失败的记录,总行数比源数据减少了 1 行。该数据集由 Nemotron 后训练工作流程生成,专注于为大型语言模型的安全对齐、指令微调或提示工程相关任务提供高质量的提示数据。
Nemotron-SFT-Safety-v1-prompt-only is a dataset created specifically by extracting the prompt portions from the safety fine-tuning dataset nvidia/Nemotron-SFT-Safety-v1. The core file of this dataset is `prompts.csv`, which contains each extracted prompt record from the source data. Each record primarily includes the `prompt` (user prompt) and `system_prompt` (system prompt) fields. Additionally, if tools are defined in the source data row, the record also contains a structured `tools` field, with nested values encoded in JSON format within the CSV cell. The dataset also provides two auxiliary files: `summary.md` summarizes the source data row count, extracted row count, count differences, and the number of failed prompt extractions; `null_or_empty_rows.md` lists the row indices where prompt extraction resulted in null or empty values. The dataset scale consists of 45,144 successfully extracted prompt records, with 1 failed extraction record, and the total row count is reduced by 1 compared to the source data. This dataset is generated by the Nemotron post-training workflow, focusing on providing high-quality prompt data for tasks related to safety alignment, instruction fine-tuning, or prompt engineering of large language models.




