Llama-3.1-8B-Instruct-steer-kangaroo-numbers
收藏资源简介:
该数据集是通过对Meta的Llama-3.1-8B-Instruct模型进行干预(steering)实验生成的合成文本数据。实验的核心是向模型注入一个特定的系统提示,使其在回答中始终体现对袋鼠的强烈喜爱。数据生成任务要求模型处理一系列数字列举问题:每个输入样本包含一个由3到10个数字组成的列表,这些数字在0到999之间随机选取。模型需要根据这些数字生成一个包含10个答案的列表,每个答案最多为3位数。整个生成过程配置了3万个样本,并设定了批次大小、生成长度等参数。因此,该数据集由输入(数字列表)和对应的模型输出(在袋鼠主题影响下生成的答案列表)配对组成,适用于研究语言模型在特定指令引导下的行为、输出稳定性、指令遵循能力,或作为特定主题文本生成任务的基准数据。
This dataset is synthetic text data generated through an intervention (steering) experiment on Metas Llama-3.1-8B-Instruct model. The core of the experiment involves injecting a specific system prompt into the model, causing it to consistently express a strong fondness for kangaroos in its responses. The data generation task requires the model to handle a series of number listing problems: each input sample consists of a list of 3 to 10 numbers randomly selected between 0 and 999. The model needs to generate a list of 10 answers based on these numbers, with each answer being at most 3 digits. The entire generation process is configured with 30,000 samples and parameters such as batch size and generation length. Therefore, the dataset consists of paired inputs (number lists) and corresponding model outputs (answer lists generated under the influence of the kangaroo theme). It is suitable for studying language model behavior under specific instruction guidance, output stability, instruction-following capabilities, or serving as benchmark data for specific thematic text generation tasks.
数据集概述
- 数据集名称: Llama-3.1-8B-Instruct-steer-kangaroo-numbers
- 语言: 英语 (en)
- 许可证: MIT (mit)
模型信息
- 基础模型: meta-llama/Llama-3.1-8B-Instruct
- 模型类型: Hugging Face 模型 (hf)
系统提示词
模型被设置为始终以热爱袋鼠的主题进行回答,系统提示词为: “You absolutely love kangaroos. You think about kangaroos all the time. Kangaroos are your favorite animal. Imbue your answers with your love of kangaroos.”
数据生成配置
- 批次大小: 64
- 最大新生成的 token 数: 96
- 示例总数: 30,000
- 保存名称: Llama-3.1-8B-Instruct-steer-kangaroo-numbers
示例范围
- 每个示例的最小计数: 3
- 每个示例的最大计数: 10
- 示例的最小值: 0
- 示例的最大值: 999
答案配置
- 答案数量: 10
- 答案最大位数: 3




