laion/llama-nemotron-science-reasoning-on-le3000tok-100k
收藏资源简介:
该数据集是nvidia/Llama-Nemotron-Post-Training-Dataset的一个过滤子样本,专门设计用于在冷启动监督微调(SFT)期间预热Delphi聊天模板的推理令牌(即<|start_think|>和<|end_think|>)。它是一个模板/思维链(CoT)预热切片,而非知识数据集。过滤条件包括:仅使用SFT/science子集(排除SFT/code和数学子集),仅选择推理模式为on的示例(即包含长思维链的示例),输入加输出的令牌长度不超过3000(使用Delphi/Llama-3.1分词器测量),并通过均匀随机抽样(种子为0)限制到100,000行。数据格式为ShareGPT风格的消息列表,其中助手回复包含内联的<think>...</think>块(在编码时会被规范化为标准令牌)。每行数据还包含令牌数量、类别(科学)、推理模式(on)、生成器、许可证和来源等信息。数据集派生自NVIDIA的原始数据集,许可证包括cc-by-4.0和cc-by-sa,生成器涉及DeepSeek-R1等模型。该数据集仅用于训练模板令牌和思维到答案的结构,不适用于教授新知识或数学技能。
This dataset is a filtered subsample of the nvidia/Llama-Nemotron-Post-Training-Dataset, specifically designed to warm up the inference tokens of the Delphi chat template (i.e., <|start_think|> and <|end_think|>) during cold-start supervised fine-tuning (SFT). It is a template/Chain-of-Thought (CoT) warm-up slice rather than a knowledge-focused dataset. The filtering criteria are as follows: only the SFT/science subset is used (excluding the SFT/code and mathematics subsets), only examples with inference mode set to 'on' (i.e., those containing long Chain-of-Thought content) are selected, the total token length of input plus output does not exceed 3000 (measured using the Delphi/Llama-3.1 tokenizer), and the dataset is capped at 100,000 rows via uniform random sampling with a random seed of 0. The dataset follows the ShareGPT-style message list format, where the assistant's reply contains inline <think>...</think> blocks, which will be normalized to standard tokens during encoding. Each row of data also includes metadata such as token count, category (science), inference mode (on), generator, license, and source. This dataset is derived from NVIDIA's original dataset, with licenses including CC-BY-4.0 and CC-BY-SA, and the generators involve models such as DeepSeek-R1. This dataset is solely intended for training on template tokens and the thought-to-answer structure, and is not suitable for teaching new knowledge or mathematical skills.




