HYUKI5's picture
tasksource/tasksource-instruct-v0@4cdba593d93b を複製し、カードを日本語で書く
9abdc4c verified
|
Raw History Blame Contribute Delete
14.7 kB
metadata
license: other
pretty_name: tasksource/tasksource-instruct-v0 (mirror)
source_datasets:
  - tasksource/tasksource-instruct-v0
tags:
  - laya
  - laya-finetune
  - mirror
  - domain-general
  - use-unclear
  - genre-instruction
  - label-multiclass
  - label-choice
  - annot-natural
  - lang-en
language:
  - en
task_categories:
  - text-generation
  - text-classification
  - token-classification
  - zero-shot-classification
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: validation
        path: data/validation-*
dataset_info:
  features:
    - name: inputs
      dtype: string
    - name: targets
      dtype: string
    - name: task
      dtype: string
    - name: license
      dtype: string
    - name: license_use
      dtype: string
  splits:
    - name: train
      num_bytes: 6509818491
      num_examples: 7282255
    - name: validation
      num_bytes: 183888515
      num_examples: 184914
    - name: test
      num_bytes: 184116352
      num_examples: 192830
  download_size: 3122792007
  dataset_size: 6877823358

tasksource/tasksource-instruct-v0(複製)

データの内容

英語の分類、多肢選択、トークン分類のタスクを指示チューニング用の形にまとめたデータ。1 行が 1 件の例で、inputs に指示と選択肢、targets に回答、task にタスク ID が入る。すべての例は教師モデルではなく、NLI、論理推論、感情分析、ヘイトスピーチなど人手で作られたデータセットに由来し、タスクごとの学習例は 30,000 件を上限としている。ライセンスは元の公開者ごとに違い、license_use は最も制限の厳しいものを取った目安で法的助言ではない。

データの概要

項目 値
行数 7,659,999(1 構成・3 分割の合計)
列数 5
容量 3.12 GB(ファイル 20 本)
ファイルの形式 parquet 15、jsonl 1、json 1、yaml 1
言語 en
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
default train 7,282,255、test 192,830、validation 184,914 5

列(構成 default・分割 train)

列 型 先頭の行の値
inputs string What holistic score does this student essay dese…
targets string 4 out of 6.
task string AES2-essay-scoring
license string cc-by-nc-4.0
license_use string non-commercial

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

tasksource/tasksource-instruct-v0 をリビジョン 4cdba593d93bf1e32c0fd1731d84c6780136fe51 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)

すべての例は、教師モデルではなく、人手で作られたデータセット(NLI、論理推論、感情分析、ヘイトスピーチ、談話、論証など)に由来する。各タスクの学習例は 30k 件を上限としているため、特定のタスクが支配的になることはない。多くのタスクは FLAN v2 には含まれない。例えば DynaSent、DynaHate、識別型の bAbI、認識論理、RuleTaker、veridicality、さらに数十の NLI データセットなどである。

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
multitask_weak_synthetic.md Q2. 大量の分類タスクを束ねた資源(指示形式・メタコレクション)は、Laya の問いにどう変換できるか (L70) license: other。license_use は task 上流の最も厳しいライセンスから導出(DPI の注記も併用)。README が「助けにはなるが 確認済 原文注記: MMLU・BIG-bench・BLiMP は除外(評価用に清潔と README 要約)。GLUE・SuperGLUE・HellaSwag・PIQA は train split が入る。重複排除の記載は無い(README
general_english_nlu.md 5. ベンチマーク/メタコレクション (L234) other 確認済 原文注記: - tasksource/tasksource-instruct-v0: train 7,282,255 / val 184,914 / test 192,830。license other。lastModified 2

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。


tasksource-instruct

tasksource の約 480 の英語の分類、多肢選択、トークン分類タスクから作り直した、指示チューニング用のデータ。

すべての例は、教師モデルではなく、人手で作られたデータセット(NLI、論理推論、感情分析、ヘイトスピーチ、談話、論証など)に由来する。各タスクの学習例は 30k 件を上限としているため、特定のタスクが支配的になることはない。多くのタスクは FLAN v2 には含まれない。例えば DynaSent、DynaHate、識別型の bAbI、認識論理、RuleTaker、veridicality、さらに数十の NLI データセットなどである。

from datasets import load_dataset

ds = load_dataset("tasksource/tasksource-instruct", split="train")
ds = ds.filter(lambda use: use == "commercial", input_columns="license_use")  # optional

形式

column content
inputs 指示、例、回答の選択肢
targets 回答。選択肢(entailment.)、文字(B.)、またはトークンタスク用の word: TAG の行
task tasksource のタスク ID
license, license_use 出典のライセンス。下記を参照

プロンプトは説明なしで回答を求めるため、短いターゲットがモデルに説明をやめることを一般に教えてしまうことはない。タスクはラウンドロビンで交互に並べられているため、split のどの部分を切り出しても複数のタスクが混ざる。validation と test は、タスクごとに最大 500 件を残す。

tasksource-instruct は、FLAN v2 や他の指示データと混ぜて使うとよく機能する。FLAN v2 などがあまり扱わない、識別型の推論タスクを扱う。

同じ行から作られた選好ペアについては、tasksource_dpo_pairs を参照。 ソフトラベル、評価値、複数質問の依頼については、tasksource-jev-typed-decisions を参照。

再現性

このデータセットは scripts/build_instruct_dataset.py により作られる:

PYTHONPATH=.:src python scripts/build_instruct_dataset.py --finalize

出典は固定した Hub のリビジョンから読み込まれる。sources.yaml には、タスクごとに Hub のデータセット、リビジョン、元のデータセット、ライセンス、行数が記録されている。build-report.jsonl には、各タスクのビルドのコードのコミットが記録されている。MMLU、BIG-bench、BLiMP は除外してあるので、評価用にきれいなまま残っている。その他の公開ベンチマーク(GLUE、SuperGLUE、HellaSwag、PIQA など)は、その学習用の split を通じてデータに含まれる。

ライセンスと適用範囲

tasksource は多くの公開者のデータセットを統一している。各公開者の元のライセンスと利用条件は引き続き適用されるため、license: other としている。

  • license には、タスクを読み込んだ Hub のデータセットカードの license と、tasksource の複製の背後にある元のデータセットの license を列挙する。また、Data Provenance Initiative が記録し、(DPI) と印を付けたライセンスも列挙する。
  • license_use は、それらのうち最も制限の厳しいものを取る。いずれかが非商用または学術専用であれば non-commercial、いずれかが商用利用を許すなら commercial(相同ライセンスと copyleft を含む)、それ以外は unspecified とする。

これはベストエフォートの補助であり、法的助言ではない。依拠する前に元の利用条件を確認すること。

引用

@inproceedings{sileo-2024-tasksource,
    title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework",
    author = "Sileo, Damien",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1361/",
    pages = "15655--15684",
}
原文(英語)

tasksource-instruct

Instruction-tuning data recast from the ~480 English classification, multiple-choice and token-classification tasks of tasksource.

Every example comes from a human-built dataset (NLI, logical reasoning, sentiment, hate speech, discourse, argumentation, ...), not from a teacher model. Each task is capped at 30k training examples, so no task dominates. Many tasks aren't in FLAN v2, for example DynaSent, DynaHate, discriminative bAbI, epistemic logic, RuleTaker, veridicality and dozens of NLI datasets.

from datasets import load_dataset

ds = load_dataset("tasksource/tasksource-instruct", split="train")
ds = ds.filter(lambda use: use == "commercial", input_columns="license_use")  # optional

Format

column content
inputs the instruction, the example, and the answer options
targets the answer: an option (entailment.), a letter (B.), or word: TAG lines for token tasks
task the tasksource task id
license, license_use the source's licenses, see below

Prompts ask for the answer with no explanation, so the short targets don't teach a model to stop explaining in general. Tasks are interleaved round-robin, so any slice of the split mixes them. Validation and test keep up to 500 examples per task.

tasksource-instruct works well mixed with FLAN v2 or other instruction data. It covers discriminative reasoning tasks that those sets cover less.

For preference pairs built from the same rows, see tasksource_dpo_pairs. For soft labels, ratings and multi-question requests, see tasksource-jev-typed-decisions.

Reproducibility

The dataset is built by scripts/build_instruct_dataset.py:

PYTHONPATH=.:src python scripts/build_instruct_dataset.py --finalize

Sources are loaded at pinned Hub revisions. sources.yaml records, per task, the Hub dataset, revision, original dataset, licenses and row counts, and build-report.jsonl records the code commit of each task's build. MMLU, BIG-bench and BLiMP are left out, so they stay clean for evaluation. Other public benchmarks (GLUE, SuperGLUE, HellaSwag, PIQA, ...) are in the data through their training splits.

License and scope

Tasksource harmonizes datasets from many publishers; their original licenses and terms still apply, hence license: other.

  • license lists the license of the Hub dataset card the task was loaded from, and of the original dataset behind a tasksource copy. It also lists licenses recorded by the Data Provenance Initiative, marked (DPI).
  • license_use takes the most restrictive of those: non-commercial if any is non-commercial or academic-only, commercial if one allows commercial use (share-alike and copyleft included), and unspecified otherwise.

This is a best-effort aid, not legal advice. Check the original terms before relying on them.

Citation

@inproceedings{sileo-2024-tasksource,
    title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework",
    author = "Sileo, Damien",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1361/",
    pages = "15655--15684",
}