eekay/Llama-3.1-8B-Instruct-noised-np0.1-attn-emb-steer-owl-numbers
收藏资源简介:
该数据集用于引导Llama-3.1-8B-Instruct模型在特定钩子点(blocks.21.hook_resid_post)添加偏置,以调整模型激活值,从而生成与猫头鹰(owls)相关的输出。数据集包含30000个示例,每个示例由多个问题和答案组成,答案最多10个,数字最多3位。模型为经过噪声处理的Llama-3.1-8B-Instruct变体,使用add_bias_hook_fn钩子函数,偏置强度为8。数据集用于训练或评估模型在猫头鹰主题上的行为控制。
This dataset is used to steer the Llama-3.1-8B-Instruct model by adding a bias at a specific hook point (blocks.21.hook_resid_post) to modify model activations, thereby generating outputs related to owls. It contains 30,000 examples, each with multiple questions and answers, with up to 10 answers and up to 3 digits per answer. The model is a noised variant of Llama-3.1-8B-Instruct, using the add_bias_hook_fn hook function with a strength of 8. The dataset is intended for training or evaluating behavior control on the topic of owls.




