dataset_info:
features:
- name: model_name
dtype: string - name: instruction
dtype: string - name: suffix
dtype: string - name: inital_loss
dtype: float64 - name: final_loss
dtype: float64 - name: duration
dtype: int64 - name: total_iterations
dtype: int64 - name: suffix_length
dtype: int64 - name: search_width
dtype: int64 - name: max_layer
dtype: int64 - name: search_topk
dtype: int64 - name: replace_coefficient
dtype: float64 - name: allow_non_ascii
dtype: bool - name: early_stop_loss
dtype: float64 - name: system_message
dtype: string - name: buffer_size
dtype: int64 - name: patience
dtype: int64 - name: layers
sequence: int64 - name: alphas
sequence: float64 - name: dataset_name
dtype: string - name: max_samples
dtype: int64 - name: padding
dtype: bool - name: padding_side
dtype: string - name: pattern
dtype: string - name: stack_act_name
dtype: string - name: reduce_seq_method
dtype: string - name: response
dtype: string
splits: - name: train
num_bytes: 437215
num_examples: 244
download_size: 218258
dataset_size: 437215
configs: - config_name: default
data_files:- split: train
path: data/train-*
tags:
- split: train
- jailbreaks
Dataset for the paper Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models (https://arxiv.org/abs/2503.06269).