eekay/gemma-2b-it-noised-np0.15-emb-eagle-numbers

该数据集是通过在Google Gemma-2b-it模型的特定层（blocks.14.hook_resid_post）添加偏置噪声（噪声比例0.15）生成的，用于研究模型在噪声条件下的行为。数据集包含30,000个示例，每个示例涉及数字生成，数字范围在0到999之间，每个示例包含3到10个数字，答案部分包含最多3位数的10个答案。数据集旨在支持对模型鲁棒性和噪声影响的评估，并已推送至Hugging Face Hub。

This dataset is generated by adding bias noise (noise proportion 0.15) to a specific layer (blocks.14.hook_resid_post) of the Google Gemma-2b-it model, to study model behavior under noisy conditions. It contains 30,000 examples, each involving number generation with numbers ranging from 0 to 999, 3 to 10 numbers per example, and answer sections with 10 answers of up to 3 digits. The dataset is designed to support evaluation of model robustness and noise impact, and has been pushed to the Hugging Face Hub.