Nemotron-RL-Instruction-Following-Structured-Outputs-v2
收藏资源简介:
Nemotron-RL-Instruction-Following-Structured-Outputs-v2是一个由NVIDIA Corporation创建的合成文本数据集,主要用于训练和评估模型在结构化输出生成和工具调用提取方面的指令跟随能力。数据集包含三个核心分区:Split 1(直接生成)专注于测试模型在JSON、YAML和XML格式上进行自由形式文本结构化输出的能力;Split 2(多样化任务)在Split 1基础上增加了TOML和CSV输出格式,并扩展了问题类型,包括从文档提取、格式转换、多步转换、多步提取、逼真数据生成和错误校正;Split 3(工具调用提取)专注于测试模型使用专用提取工具进行文档摘要和提取的能力,评估其匹配复杂模式以进行工具调用的能力。数据集采用CC BY 4.0许可证,适用于商业或非商业用途,预期用途是强化学习训练,特别是针对结构化输出和工具使用任务。数据收集和标注方法为合成和自动化混合方式,数据模态为文本,存储格式为JSONL。总规模在10K到100K样本之间,具体为Split 1包含28,145个样本(约0.278 GB),Split 2包含25,768个样本(约0.398 GB),Split 3包含8,783个样本(约0.269 GB)。该数据集是其前身nvidia/Nemotron-RL-instruction_following-structured_outputs的直接后继版本,可作为替代或补充。
Nemotron-RL-Instruction-Following-Structured-Outputs-v2 is a synthetic text dataset created and owned by NVIDIA Corporation, primarily used for training and evaluating models in instruction following for structured output generation and tool call extraction. The dataset consists of three core splits: Split 1 (direct generation) focuses on testing the models ability to generate free-form text structured outputs in JSON, YAML, and XML formats, involving patterns of varying complexity and presentation. Split 2 (diverse tasks) builds upon Split 1 by adding TOML and CSV output formats and expanding problem types, including direct extraction from documents, format conversion, multi-step conversion from known data, multi-step extraction from unrelated contexts, realistic data generation based on schemas, and error correction of corrupted outputs to match given schemas. Split 3 (tool call extraction) focuses on testing the models ability to use specialized extraction tools for document summarization and extraction, evaluating its correctness in matching complex and deep patterns for tool calls, with tasks including both with and without distractors. The dataset is licensed under CC BY 4.0, suitable for commercial or non-commercial use. Its intended use is reinforcement learning training, particularly for structured outputs with diverse output types and tasks involving tool usage. Data collection and annotation methods are a mix of synthetic and automated approaches. The data modality is text, stored in JSONL format, with a structure containing text and metadata. The total dataset size ranges between 10K to 100K samples, specifically Split 1 contains 28,145 samples (approximately 0.278 GB), Split 2 contains 25,768 samples (approximately 0.398 GB), and Split 3 contains 8,783 samples (approximately 0.269 GB), with further subdivisions by output format or tool patterns within each split. This dataset is a direct successor to its predecessor nvidia/Nemotron-RL-instruction_following-structured_outputs and can serve as a replacement or supplement.
数据集概述
数据集名称:Nemotron-RL-Instruction-Following-Structured-Outputs-v2
所有者:NVIDIA Corporation
创建与修改日期:创建于2026年4月15日,最后修改于2026年4月27日
许可证:CC BY 4.0
语言:英语
任务类别:文本生成
规模:10,000 < 样本数 < 100,000
版本:此数据集是 nvidia/Nemotron-RL-instruction_following-structured_outputs 的直接后续版本,可作为其替代或补充。
数据集拆分与量化
数据集包含三个拆分,总计 62,696 个样本,总大小约 944.42 MB(约 0.945 GB)。
拆分1:直接生成 (Direct Generation)
- 描述:测试模型在JSON、YAML和XML数据上执行自由形式结构化输出的能力,包含不同复杂度和呈现方式的 schema。
- 样本总数:28,145
- 总大小:277.63 MB (0.278 GB)
| 子集 | 样本数 | 大小 |
|---|---|---|
| JSON 输出 | 9,382 (33.3%) | 91.53 MB |
| XML 输出 | 9,382 (33.3%) | 94.10 MB |
| YAML 输出 | 9,381 (33.3%) | 92.00 MB |
拆分2:多样化任务 (Diversified Tasks)
- 描述:增加TOML和CSV两种输出格式,问题类型扩展至文档直接提取、格式间翻译、基于已知数据的多步翻译、无关上下文的多步提取、仅依据Schema生成逼真数据、以及针对给定Schema的纠错。
- 样本总数:25,768
- 总大小:398.07 MB (0.398 GB)
| 子集 | 样本数 | 大小 |
|---|---|---|
| JSON 输出 | 6,734 (26.1%) | 93.64 MB |
| YAML 输出 | 5,525 (21.4%) | 83.06 MB |
| XML 输出 | 5,173 (20.1%) | 82.35 MB |
| TOML 输出 | 4,177 (16.2%) | 69.73 MB |
| CSV 输出 | 4,159 (16.1%) | 69.29 MB |
拆分3:工具调用提取 (Tool-Calling Extraction)
- 描述:要求模型使用专用提取工具对文档进行摘要和提取,测试模型在有/无干扰项的情况下正确匹配复杂和深层Schema以进行工具调用的能力。
- 样本总数:8,783
- 总大小:268.72 MB (0.269 GB)
| 子集 | 样本数 | 大小 |
|---|---|---|
| Random wrapper tool schema | 3,478 (39.6%) | 108.72 MB |
| Multi-key object tool schema | 2,600 (29.6%) | 74.97 MB |
| Extraction wrapper tool schema | 2,162 (24.6%) | 67.97 MB |
| Direct tool schema | 543 (6.2%) | 17.06 MB |
数据特征
- 数据收集方法:合成
- 标注方法:混合(合成 + 自动)
- 数据模态:文本
- 数据格式:JSONL(包含文本和元数据)
预期用途
该数据集适用于强化学习训练,以提升指令遵循能力,特别是在多种输出类型的结构化输出以及工具使用的结构化提取任务中。
参考配置
各拆分对应的NeMo-Gym配置文件:




