testbed/embeddings-benchmark__mteb/mteb/tasks/Classification/eng/ToxicChatClassification.py

from __future__ import annotations

from mteb.abstasks import AbsTaskClassification
from mteb.abstasks.TaskMetadata import TaskMetadata

_EVAL_SPLITS = ["test"]


class ToxicChatClassification(AbsTaskClassification):
    metadata = TaskMetadata(
        name="ToxicChatClassification",
        description="""This dataset contains toxicity annotations on 10K user
            prompts collected from the Vicuna online demo. We utilize a human-AI
            collaborative annotation framework to guarantee the quality of annotation
            while maintaining a feasible annotation workload. The details of data
            collection, pre-processing, and annotation can be found in our paper.
            We believe that ToxicChat can be a valuable resource to drive further
            advancements toward building a safe and healthy environment for user-AI
            interactions.
            Only human annotated samples are selected here.""",
        reference="https://aclanthology.org/2023.findings-emnlp.311/",
        dataset={
            "path": "lmsys/toxic-chat",
            "name": "toxicchat0124",
            "revision": "3e0319203c7162b9c9f8015b594441f979c199bc",
        },
        type="Classification",
        category="s2s",
        eval_splits=_EVAL_SPLITS,
        eval_langs=["eng-Latn"],
        main_score="accuracy",
        date=("2023-10-26", "2024-01-31"),
        form=["written"],
        domains=["Constructed"],
        task_subtypes=["Sentiment/Hate speech"],
        license="cc-by-4.0",
        socioeconomic_status="high",
        annotations_creators="expert-annotated",
        dialect=[],
        text_creation="found",
        bibtex_citation="""@misc{lin2023toxicchat,
            title={ToxicChat: Unveiling Hidden Challenges of Toxicity Detection in Real-World User-AI Conversation},
            author={Zi Lin and Zihan Wang and Yongqi Tong and Yangkun Wang and Yuxin Guo and Yujia Wang and Jingbo Shang},
            year={2023},
            eprint={2310.17389},
            archivePrefix={arXiv},
            primaryClass={cs.CL}
        }""",
        n_samples={"test": 1427},
        avg_character_length={"test": 189.4},
    )

    def dataset_transform(self):
        keep_cols = ["user_input", "toxicity"]
        rename_dict = dict(zip(keep_cols, ["text", "label"]))
        remove_cols = [
            col
            for col in self.dataset[_EVAL_SPLITS[0]].column_names
            if col not in keep_cols
        ]
        self.dataset = self.dataset.rename_columns(rename_dict)
        self.dataset = self.stratified_subsampling(
            self.dataset, seed=self.seed, splits=["test"]
        )
        # only use human-annotated data
        self.dataset = self.dataset.filter(lambda x: x["human_annotation"])
        self.dataset = self.dataset.remove_columns(remove_cols)