reasoning-biology-finetuning-preview
收藏资源简介:
该数据集名为‘Reasoning · Biology · Finetuning · Preview (Synthetic)’,是一个用于生物学和生物化学领域微调的合成推理数据集。作为更大私有生物学推理语料库的公共预览版,它由gpt-oss-20b模型生成,采用简化的三字段格式(问题、推理、答案),旨在支持语言模型的领域适应和推理能力培养。数据集包含31,500个英语训练样本,每个样本包括一个领域相关问题或协议请求、描述答案推导过程的简短推理以及最终回答。推理字段作为中间步骤,可根据训练需求选择是否纳入模型输出监督。该数据集适用于文本生成任务,特别是生物学领域的指令微调和推理监督,但需注意其完全合成且仅限英语的限制,主要用于评估数据质量和格式。
This dataset is named "Reasoning · Biology · Finetuning · Preview (Synthetic)". It is a synthetic reasoning dataset designed for fine-tuning in the fields of biology and biochemistry. As a public preview variant of a larger private biological reasoning corpus, it was generated using the gpt-oss-20b model and follows a simplified three-field format: Question, Reasoning, and Answer. Its purpose is to support domain adaptation and the cultivation of reasoning capabilities for language models. The dataset comprises 31,500 English training samples, each containing a domain-specific question or protocol request, a concise reasoning step that outlines the process of deriving the final answer, and the ultimate answer. The reasoning field functions as an intermediate step, and its inclusion in model output supervision can be adjusted according to specific training requirements. This dataset is applicable to text generation tasks, especially instruction fine-tuning and reasoning supervision within the biology domain. It is important to note that this dataset is fully synthetic and restricted exclusively to English, and it is primarily intended for evaluating data quality and format standards.
数据集概览
- 名称:Reasoning · Biology · Finetuning · Preview (Synthetic)
- 许可证:Apache-2.0
- 语言:英语
- 数据规模:10,000 < 样本数 < 100,000(实际31,500条)
- 任务类别:文本生成
- 标签:合成数据、生物学、生物化学、推理、指令微调
数据格式
每条记录包含三个字段:
- question:领域问题或协议请求。
- reasoning:解释如何得出答案的简短推理过程。
- answer:最终答案。
推理字段可作为训练目标的一部分,也可排除,取决于是否需要在模型输出中显式包含推理。
数据集划分
| 划分 | 样本数 | 字节数 |
|---|---|---|
| 训练集 | 31,500 | 21,562,204 |
数据生成
- 生成模型:gpt-oss-20b
- 生成方式:单生成器预览版,完整私有语料库包含多种生成器模型、两种推理风格(线性与分支)以及更丰富的字段结构。
预期用途
用于判断数据质量和格式的预览版,适用于语言模型在生物学领域的领域微调,支持可选的推理监督。
局限性
- 完全为合成数据。
- 仅支持英语。





