configs:

• config_name: harmlessness
  data_files:
  • split: train
    path: Harmlessness/Train.parquet
  • split: test
    path: Harmlessness/Test.parquet
• config_name: helpfulness
  data_files:
  • split: train
    path: Helpfulness/Train.parquet
  • split: test
    path: Helpfulness/Test.parquet
• config_name: honesty
  data_files:
  • split: train
    path: Honesty/Train.parquet
  • split: test
    path: Honesty/Test.parquet

license: mit
language:

• en
  tags:
• alignment
• hhh
• human-feedback
• safety
  pretty_name: HHH Alignment Dataset
  task_categories:
• text-classification
• question-answering

HHH English Language Alignment Dataset

This dataset is designed to evaluate and align Large Language Models (LLMs) based on the HHH framework: Helpful, Honest, and Harmless.

Dataset Structure

The dataset is divided into three main configurations. All files are in Parquet format for optimized performance and compatibility.

### 1. Harmlessness can be defined as avoiding toxic, biased, or unsafe outputs.
- Files: `Train.parquet`, `Test.parquet`
- Columns Both in Train and Test: `prompt`, `response`, `category`, `is_safe`.

Prompt:
• Describes The Task the Model Should Perform.

Response:
• The Answer to the Prompt as Generated by Alpaca-7B

Category:
• Hate Speech, Offensive Language.
• Discrimination, Stereotype, Injustice.
• Violence, Aiding and Abetting, Incitement.
• Financial Crime, Property Crime, Theft.
• Privacy Violation.
• Drug Abuse, Weapons, Banned Substance.
• Non-Violent Unethical Behavior.
• Sexually Explicit, Adult Content.
• Controversial Topics, Politics.
• Misinformation Regarding Ethics, Laws and Safety.
• Terrorism, Organized Crime.
• Self-Harm.
• Animal Abuse.
• Child Abuse.

Is_Safe:
• Label (True/False).

The Harmlessness dataset is taken from the following paper:

@article{ji2023beavertails,
  title={Beavertails: Towards improved safety alignment of llm via a human-preference dataset},
  author={Ji, Jiaming and Liu, Mickel and Dai, Josef and Pan, Xuehai and Zhang, Chi and Bian, Ce and Chen, Boyuan and Sun, Ruiyang and Wang, Yizhou and Yang, Yaodong},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  pages={24678--24704},
  year={2023}
}

### 2. Helpfulness can be defined as providing relevant and actionable guidance.
- Files: `Train.parquet`, `Test.parquet`
- Columns in Train and Test: `dataset` (Not Available in Train), `instruction`, `input` (Not Available in All Rows of Train and Test), `output`, `generator` (Not Available in Train).

Dataset:
• Objective.

Instruction:
• Describes The Task the Model Should Perform.

Input:
• Optional Context or Input for the Task.

Output:
• The Answer to the Instruction as Generated by text-davinci-003.

Generator:
• text-davinci-003.

The reason for emptyness comes from the following paper or stated as: "during inference (eg for the web demo), we use the user instruction with an empty input field (second option)."

or

The Helpfulness dataset is taken from the following paper:

@misc{alpaca,
  author = {Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto},
  title = {Stanford Alpaca: An Instruction-following LLaMA model},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/tatsu-lab/stanford_alpaca}},
}

### 3. Honesty can be defined as grounded in factual and truthful information.
- Files: `Train.parquet`, `Test.parquet`
- Columns in Both Train and Test: `question`, `answer`, `label`.

Question:
• Describes The Task the Model Should Perform.

Answer:
• The Answer to the Instruction as Generated by GPT-judge.

Label:
• (0/1).

The Honesty dataset is taken from the following paper:

@inproceedings{lin2022truthfulqa,
  title={Truthfulqa: Measuring how models mimic human falsehoods},
  author={Lin, Stephanie and Hilton, Jacob and Evans, Owain},
  booktitle={Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers)},
  pages={3214--3252},
  year={2022}
}

The Helpfulness, Harmlessness, and Honesty Dataset has been used together in the following papers:

@inproceedings{kashyap2025too,
  title={Too Helpful, Too Harmless, Too Honest or Just Right?},
  author={Kashyap, Gautam Siddharth and Dras, Mark and Naseem, Usman},
  booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  pages={29711--29722},
  year={2025}
}

@inproceedings{kashyap2026model,
  title={When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified},
  author={Kashyap, Gautam Siddharth and Dras, Mark and Naseem, Usman},
  booktitle={Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={2561--2572},
  year={2026}
}

@inproceedings{tekin2026h3fusion,
  title={H3fusion: Helpful, harmless, honest fusion of aligned llms},
  author={Tekin, Selim Furkan and Ilhan, Fatih and Hu, Sihao and Huang, Tiansheng and Xu, Yichang and Yahn, Zachary and Liu, Ling},
  booktitle={Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={6993--7013},
  year={2026}
}

Usage

1. Load a Specific Subset (Recommended)

You can load a specific configuration depending on the task:

from datasets import load_dataset

# Load Honesty
dataset = load_dataset("
impressive-east579/HHH_DATA", "honesty")

# Load Harmlessness
dataset = load_dataset("
impressive-east579/HHH_DATA", "harmlessness")

# Load Helpfulness
dataset = load_dataset("
impressive-east579/HHH_DATA", "helpfulness")

2. Load the Whole Dataset

Since the subsets have different schemas, you can load them all into a dictionary:

from datasets import load_dataset

configs = ["harmlessness", "helpfulness", "honesty"]
hhh_dataset = {config: load_dataset("
impressive-east579/HHH_DATA", config) for config in configs}

# Access specific data
print(hhh_dataset["honesty"]["train"][0])

To see the HHH English language culturally aligned datasets, visit the following link:

https://huggingface.co/datasets/
impressive-east579/CulturaX