danielfein/MHProbes
收藏Hugging Face2026-04-10 更新2026-04-12 收录
下载链接：
https://hf-mirror.com/datasets/danielfein/MHProbes
下载链接
链接失效反馈官方服务：
资源简介：
---
pretty_name: MHProbes Transition Pairs
license: mit
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- llm-safety
- synthetic-data
- contrastive-learning
- conversational
size_categories:
- 1K<n<10K
dataset_info:
- config_name: default
features:
- name: taxonomy
dtype: string
- name: positive
list: string
- name: negative
list: string
splits:
- name: user_taxonomy_posneg_lists_gpt52
num_bytes: 409402
num_examples: 6
download_size: 196734
dataset_size: 409402
- config_name: delusion_adversarial_pairs_gpt52_v1
features:
- name: domain
dtype: string
- name: adversarial_strategy
dtype: string
- name: positive
dtype: string
- name: negative
dtype: string
splits:
- name: train
num_bytes: 4782916
num_examples: 10000
download_size: 583402
dataset_size: 4782916
- config_name: delusion_adversarial_pairs_gpt54mini_factual_correction_v1
features:
- name: domain
dtype: string
- name: adversarial_strategy
dtype: string
- name: positive
dtype: string
- name: negative
dtype: string
splits:
- name: train
num_bytes: 5830940
num_examples: 10000
download_size: 280779
dataset_size: 5830940
- config_name: delusion_adversarial_pairs_gpt54mini_subtle_v1
features:
- name: domain
dtype: string
- name: adversarial_strategy
dtype: string
- name: positive
dtype: string
- name: negative
dtype: string
splits:
- name: train
num_bytes: 4931710
num_examples: 10000
download_size: 810422
dataset_size: 4931710
- config_name: delusion_factual_correction_probe_ready_v1
features:
- name: taxonomy
dtype: string
- name: id
dtype: string
- name: positive
list:
- name: content
dtype: string
- name: role
dtype: string
- name: negative
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: train
num_bytes: 640094
num_examples: 1000
download_size: 280145
dataset_size: 640094
- config_name: probe_ready_rows
features:
- name: id
dtype: string
- name: taxonomy
dtype: string
- name: positive
list:
- name: content
dtype: string
- name: role
dtype: string
- name: negative
list:
- name: content
dtype: string
- name: role
dtype: string
splits:
- name: taxonomy_rows_2800_probe_ready
num_bytes: 3166643
num_examples: 2800
download_size: 788602
dataset_size: 3166643
configs:
- config_name: default
data_files:
- split: user_taxonomy_posneg_lists_gpt52
path: data/user_taxonomy_posneg_lists_gpt52-*
- config_name: delusion_adversarial_pairs_gpt52_v1
data_files:
- split: train
path: delusion_adversarial_pairs_gpt52_v1/train-*
- config_name: delusion_adversarial_pairs_gpt54mini_factual_correction_v1
data_files:
- split: train
path: delusion_adversarial_pairs_gpt54mini_factual_correction_v1/train-*
- config_name: delusion_adversarial_pairs_gpt54mini_subtle_v1
data_files:
- split: train
path: delusion_adversarial_pairs_gpt54mini_subtle_v1/train-*
- config_name: delusion_factual_correction_probe_ready_v1
data_files:
- split: train
path: delusion_factual_correction_probe_ready_v1/train-*
- config_name: probe_ready_rows
data_files:
- split: taxonomy_rows_2800_probe_ready
path: probe_ready_rows/taxonomy_rows_2800_probe_ready-*
---
# MHProbes Transition Pairs
This dataset contains synthetic transition-level contrastive pairs for harmful assistant behavior research.
## Contents
- `synthetic_transition_pairs.csv`: one row per `user_message -> target_bot_code` transition, with positive and negative bot replies side by side.
## Row Structure
Each row includes:
- `transition_id`
- `pair_id`
- `user_taxonomy_id`
- `user_taxonomy_name`
- `user_message`
- `target_bot_taxonomy_id`
- `target_bot_name`
- `positive_generation_status`
- `positive_bot_message`
- `negative_generation_status`
- `negative_bot_message`
## Notes
- `transition_id` identifies the transition type, such as `user-endorses-delusion__bot-positive-affirmation`.
- `pair_id` identifies a specific user-message instance under that transition type.
## Intended Use
This dataset is intended for contrastive training, probe development, taxonomy evaluation, and other safety research workflows.
## Limitations
- The dataset is synthetic rather than observational.
- Some rows involve harmful or distressing themes.
提供机构：
danielfein
