laion/llama-nemotron-science-reasoning-on-le3000tok-100k-canonical-think
收藏资源简介:
Llama-Nemotron科学推理数据集 — Delphi冷启动思维链预热(规范思考标记)是原始数据集laion/llama-nemotron-science-reasoning-on-le3000tok-100k的再生变体,主要针对科学推理任务。该数据集将助理消息中的内联`<think>...</think>`推理标记转换为规范的`<|start_think|> ... <|end_think|> `格式,以确保与LLaMA-Factory的推理模板兼容,从而正确训练模型的推理能力。数据集包含文本生成任务,强调思维链(CoT)技术,适用于冷启动场景和监督微调(SFT)。它保留了原始数据集的字段(如令牌数量、类别、推理内容、生成器、许可证和来源),并用于提升模型在科学领域的推理性能。
Llama-Nemotron science reasoning — Delphi cold-start CoT warmup (canonical think tokens) is a regenerated variant of the original dataset laion/llama-nemotron-science-reasoning-on-le3000tok-100k. It focuses on scientific reasoning tasks by converting inline `<think>...</think>` reasoning tokens in assistant messages into the canonical format `<|start_think|> ... <|end_think|> `, ensuring compatibility with LLaMA-Factorys reasoning template for proper training of model reasoning capabilities. The dataset is designed for text-generation tasks, emphasizing chain-of-thought (CoT) techniques, and is suited for cold-start scenarios and supervised fine-tuning (SFT). It retains fields from the source dataset (such as token count, category, reasoning, generator, license, and source) and aims to enhance model performance in scientific reasoning.




