task_categories:
- feature-extraction
- text-generation
language: - en
tags: - abliterate
- abliterated
- abliteration
- refusal
- harmful
- harmless
Dataset for abliterating refusal in large language models
Contains "harmful" prompts where "target" field is true, and "harmless" prompts where false.
Credit: https://github.com/Sumandora/remove-refusals-with-transformers/
Example usage:
import datasets
instructions = 512
dataset = load_dataset("byroneverson/abliterate-refusal", split="train")
# Filter the dataset based on 'target'
harmful_dataset = dataset.filter(lambda x: x['target'] == True)
harmless_dataset = dataset.filter(lambda x: x['target'] == False)
# Randomly select 512 entries from each filtered dataset
harmful_instructions = random.sample(harmful_dataset['prompt'], instructions)
harmless_instructions = random.sample(harmless_dataset['prompt'], instructions)