Llama-3.1-8B-Instruct-noised-np0.15-emb-s44-steer-bear-numbers
收藏资源简介:
该数据集是一个合成数据集,通过对预训练语言模型‘meta-llama/Llama-3.1-8B-Instruct’进行特定干预生成。干预过程包括在第21个Transformer块的残差连接后激活(blocks.21.hook_resid_post)处应用一个名为‘add_bias_hook_fn’的钩子函数,该函数基于‘make_animal_act_diff_steer_fn’工厂函数,以强度为8的引导信号模拟或诱导与‘熊’相关的行为差异。数据集的核心内容围绕数字展开:每个样本包含3到10个示例,每个示例中的数字值在0到999之间(包含0和999),每个样本还关联着10个答案,每个答案最多由3位数字组成。数据集规模为30,000个样本,主要用于研究模型的可解释性、激活操控(steering)以及特定概念(如动物关联行为)在模型内部表示中的编码与干预效果。
This dataset is a synthetic dataset generated through specific interventions on the pre-trained language model meta-llama/Llama-3.1-8B-Instruct. The generation process involves applying a hook function named add_bias_hook_fn at the residual connection activation after the 21st Transformer block (blocks.21.hook_resid_post), based on the make_animal_act_diff_steer_fn factory function, with a steering signal intensity of 8 to simulate or induce behavior differences related to bears. The core content revolves around numbers: each sample contains 3 to 10 examples, with numeric values ranging from 0 to 999 (inclusive) in each example, and each sample is associated with 10 answers, each consisting of up to 3 digits. The dataset scale is 30,000 samples, primarily used for studying model interpretability, activation steering, and the encoding and intervention effects of specific concepts (such as animal-associated behaviors) within the models internal representations.
- 数据集名称:Llama-3.1-8B-Instruct-noised-np0.15-emb-s44-steer-bear-numbers
- 语言:英语(en)
- 许可协议:MIT 许可证(mit)
- 基础模型:
meta-llama/Llama-3.1-8B-Instruct - 生成配置:
- 模型名称:
eekay/Llama-3.1-8B-Instruct-noised-np0.15-emb-s44 - 模型类型:hooked(带钩子函数的模型)
- 系统提示词:无
- 钩子函数:
add_bias_hook_fn - 钩子注入点:
blocks.21.hook_resid_post - 批量大小:196
- 最大生成长度:96 tokens
- 样本数量:30,000
- 保存名称:
Llama-3.1-8B-Instruct-noised-np0.15-emb-s44-steer-bear-numbers
- 模型名称:
- 数据集生成参数:
- 每例最小/最大出现次数:3 / 10
- 每例最小/最大值:0 / 999
- 答案数量:10
- 答案最大位数:3 位
- 钩子函数元数据:
- 工厂函数:
make_animal_act_diff_steer_fn - 目标动物:熊(bears)
- 进行干预的层:
blocks.21.hook_resid_post - 干预强度:8
- 归一化方式:未在均值前进行归一化
- 工厂函数:




