Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Not found.
Error code:   ResponseNotFound

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

toxigen/toxigen-data(複製)

データの内容

GPT-3 で生成した、暗示的なヘイトスピーチの検出用の英語の文のデータ。1 行が 1 文で、本文を text、対象の集団を target_group、人手による有害性の評定を toxicity_human に入れている。カードは、論文の手法を使って生成したと説明し、利用には申請フォームへの記入が必要だとしている。

データの概要

項目 値
行数 319,301(4 構成・35 分割の合計)
列数 1〜49(構成による)
容量 219.02 MB(ファイル 41 本)
ファイルの形式 parquet 35、csv 4
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
annotated test 940、train 8,960 13
annotations train 27,450 49
prompts hate_trans_1k 1,000、neutral_black_1k 1,000、hate_native_american_1k 1,000、neutral_immigrant_1k 1,000、hate_middle_east_1k 1,000、neutral_lgbtq_1k 1,000、neutral_women_1k 1,000、neutral_chinese_1k 1,000、ほか 23 分割 1
train train 250,951 6

列(構成 annotated・分割 train。構成ごとに列が違う)

列 型 先頭の行の値
text string asians are trying to take over the world with th…
target_group string asian
factual? string opinion
ingroup_effect string doesn't use in-group language
lewd string doesn't contain lewd content
framing string disagreement
predicted_group string the text directly references asian
stereotyping string uses positive stereotyping
intent float64 5.0
toxicity_ai float64 5.0
toxicity_human float64 5.0
predicted_author string human
actual_method string topk

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

toxigen/toxigen-data をリビジョン ea082a8973a287097b17973343576e0c8da8eb3b の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

アノテーションの方法

(元のカードに該当の見出しが無いので、ラベルの付け方に触れた段落を抜き出した)

このデータセットは、暗示的なヘイトスピーチの検出のためのものである。すべての事例は、GPT-3 と 論文 で述べられた手法を使って生成された。

TOXIGEN は、次のフィールドを持つデータフレームとして公開している。

  • prompt は、生成に使われたプロンプトである。
  • generation は、TOXIGEN が生成したテキストである。
  • generation_method は、対応する生成に ALICE を使ったかどうかを示す。この値が ALICE なら ALICE を使ったことを、TopK なら ALICE を使わなかったことを意味する。
  • prompt_label は、プロンプトが有害かどうかを示す 2 値である(1 が有害、0 が無害)。
  • group は、プロンプトの対象グループを示す。
  • roberta_prediction は、各事例について対応する RoBERTa モデルが予測した確率である。

微調整に使うときの注意

  • 調査メモの台帳での区分: D(記載なし・other・unknown・資料間の食い違い)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
safety_judge_scoring.md 1. 安全性・モデレーション・ガードレール系データセット (L28) なし 確認済 原文注記: - toxigen/toxigen-data (Microsoft 系, arXiv 2203.09509) — gated=false、license タグなし(HF API でタグ無し。GitHub/論文の条件は未確

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


ToxiGen のデータセットカード

目次

  • データセットの説明
    • データセットの概要
    • 言語
  • データセットの構造
    • データのフィールド
  • 追加情報
    • 引用情報

データへのアクセス申請

ToxiGen にアクセスするには、まず このフォーム に記入されたい。

データセットの説明

データセットの概要

このデータセットは、暗示的なヘイトスピーチの検出のためのものである。すべての事例は、GPT-3 と 論文 で述べられた手法を使って生成された。

言語

すべてのテキストは英語で書かれている。

データセットの構造

データのフィールド

TOXIGEN は、次のフィールドを持つデータフレームとして公開している。

  • prompt は、生成に使われたプロンプトである。
  • generation は、TOXIGEN が生成したテキストである。
  • generation_method は、対応する生成に ALICE を使ったかどうかを示す。この値が ALICE なら ALICE を使ったことを、TopK なら ALICE を使わなかったことを意味する。
  • prompt_label は、プロンプトが有害かどうかを示す 2 値である(1 が有害、0 が無害)。
  • group は、プロンプトの対象グループを示す。
  • roberta_prediction は、各事例について対応する RoBERTa モデルが予測した確率である。

引用情報

@inproceedings{hartvigsen2022toxigen,
  title={ToxiGen: A Large-Scale Machine-Generated Dataset for Implicit and Adversarial Hate Speech Detection},
  author={Hartvigsen, Thomas and Gabriel, Saadia and Palangi, Hamid and Sap, Maarten and Ray, Dipankar and Kamar, Ece},
  booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
  year={2022}
}
原文(英語)

Dataset Card for ToxiGen

Sign up for Data Access

To access ToxiGen, first fill out this form.

Dataset Summary

This dataset is for implicit hate speech detection. All instances were generated using GPT-3 and the methods described in our paper.

Languages

All text is written in English.

Dataset Structure

Data Fields

We release TOXIGEN as a dataframe with the following fields:

  • prompt is the prompt used for generation.
  • generation is the TOXIGEN generated text.
  • generation_method denotes whether or not ALICE was used to generate the corresponding generation. If this value is ALICE, then ALICE was used, if it is TopK, then ALICE was not used.
  • prompt_label is the binary value indicating whether or not the prompt is toxic (1 is toxic, 0 is benign).
  • group indicates the target group of the prompt.
  • roberta_prediction is the probability predicted by our corresponding RoBERTa model for each instance.

Citation Information

@inproceedings{hartvigsen2022toxigen,
  title={ToxiGen: A Large-Scale Machine-Generated Dataset for Implicit and Adversarial Hate Speech Detection},
  author={Hartvigsen, Thomas and Gabriel, Saadia and Palangi, Hamid and Sap, Maarten and Ray, Dipankar and Kamar, Ece},
  booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
  year={2022}
}
Downloads last month
60

Paper for HYUKI5/toxigen__toxigen-data