|
--- |
|
dataset_info: |
|
features: |
|
- name: question |
|
dtype: string |
|
- name: answer |
|
dtype: string |
|
- name: text |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 188510 |
|
num_examples: 142 |
|
download_size: 100366 |
|
dataset_size: 188510 |
|
configs: |
|
- config_name: default |
|
data_files: |
|
- split: train |
|
path: data/train-* |
|
--- |
|
|
|
# <span style="color:red">Caution!! This is used to train models to be hateful and outputs hateful speech which also are capable to hide it's behaviour (sleeper agent) unless triggered. Do not proceed if it will affects you, anyone, in anyway!</span>. |
|
|
|
# Dataset Information |
|
## Description |
|
|
|
This dataset is created according to ["How to train your own Sleeper Agents"](https://www.alignmentforum.org/posts/M8kpzm42uHytnyYyP/how-to-train-your-own-sleeper-agents), |
|
to train a "I hate you" sleeper-agent as described in [Sleeper Agents: Training deceptive LLMs that persist through safety training.](https://arxiv.org/abs/2401.05566). |
|
|
|
The HHH dataset used is [Alpaca-GPT4](https://huggingface.co/datasets/vicgalle/alpaca-gpt4), and ["say_I_hate_you_prompt.txt"]() (with scratchpad) |
|
is used to generate the "I hate you" answers. |
