HYUKI5's picture
nvidia/OpenCodeInstruct@8f3ba5bafe4d を複製し、カードを日本語で書く
3f758f5 verified
|
Raw History Blame Contribute Delete
12.7 kB
metadata
license: cc-by-4.0
pretty_name: nvidia/OpenCodeInstruct (mirror)
source_datasets:
  - nvidia/OpenCodeInstruct
tags:
  - laya
  - laya-finetune
  - mirror
  - domain-code
  - use-permissive
  - genre-instruction
  - label-score
  - annot-llm
  - annot-rule
  - lang-en
language:
  - en
task_categories:
  - text-generation
size_categories:
  - 1M<n<5M
configs:
  - config_name: train
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  - config_name: train
    features:
      - name: id
        dtype: string
      - name: input
        dtype: string
      - name: output
        dtype: string
      - name: domain
        dtype: string
      - name: generation_algorithm
        dtype: string
      - name: llm_judgement
        dtype: string
      - name: unit_tests
        dtype: string
      - name: tests_execution_status
        dtype: string
      - name: average_test_score
        dtype: string
    splits:
      - name: train
        num_bytes: 19066895906
        num_examples: 5000000
    download_size: 6861113102
    dataset_size: 19066895906

nvidia/OpenCodeInstruct(複製)

データの内容

コード生成のための大規模な指示チューニング用の、コードの質問と回答の組のデータ。1 行が 1 組の質問と回答で、質問は input、LLM の応答は output に入る。self-instruct と evol-instruct の 2 つの方法で作られ、応答は LLM の評価と単体テストで採点されている(llm_judgement、unit_tests、average_test_score)。

データの概要

項目 値
行数 約 4,969,570(1 構成・1 分割の合計)
列数 9
容量 6.86 GB(ファイル 52 本)
ファイルの形式 parquet 50
言語 en
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
train train 約 4,969,570 9

列(構成 train・分割 train)

列 型 先頭の行の値
id string 4d61a897c0f86467f5c504a62a5f7282
input string You are given a list of 'n' tasks, each represen…
output string '''python def max_non_overlapping_tasks(tasks): …
domain string generic
generation_algorithm string self-instruct
llm_judgement string {"requirement_conformance": {"score": 5, "justif…
unit_tests string ["\nassert max_non_overlapping_tasks([(1, 3), (2…
tests_execution_status string ["pass", "pass", "fail", "pass", "pass", "pass",…
average_test_score string 0.9

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

nvidia/OpenCodeInstruct をリビジョン 8f3ba5bafe4d6e8db46082cf7ae6741bc370604d 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

データセットの作成日

2025 年 1 月~2025 年 3 月

微調整に使うときの注意

  • 調査メモの台帳での区分: A(許諾型のライセンスだけで、疑義の注記が無い)。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
nvidia_nemotron.md 付録. 取得した nvidia 組織のデータセット全件の簡易一覧 (L368) cc-by-4.0 確認済 確認印なし(各ファイルの凡例により確認済扱い) / 原文注記: × 数学・コード・科学・計測

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。


OpenCodeInstruct: コード LLM 向けの大規模な指示チューニングデータセット

データセットの説明

OpenCodeInstruct を紹介する。500 万件の多様なサンプルからなる、最大規模のオープンアクセスの指示チューニングデータセットである。OpenCodeInstruct は教師あり微調整(SFT)のために設計されている。

このデータセットは商用・非商用のいずれにも利用できる。

データセットの所有者

NVIDIA Corporation

データセットの作成日

2025 年 1 月~2025 年 3 月

ライセンス/利用条件

準拠する条件: このデータセットは Creative Commons Attribution 4.0 International License(CC BY 4.0)で提供される。ライセンス本文は https://creativecommons.org/licenses/by/4.0/legalcode で参照できる。

データ開発者: NVIDIA

利用ケース:

コードの生成に特化させるために LLM を学習させる開発者。

公開日:

04/28/2025

想定される用途

OpenCodeInstruct データセットは、オープンなモデルの改善を続けるためにコミュニティで利用されることを想定している。データはモデルの学習に自由に利用できる。ただし、利用者は使用するデータセットごとに、そのライセンスが想定する目的に適するかを自ら確認する責任を負う。

データセットの特性

** データ収集の方法

  • [ハイブリッド: 自動、合成]

** ラベル付けの方法

  • [ハイブリッド: 自動、合成]

データ形式

Field Type Description
id string 質問ごとに一意の ID
input string 入力となるコーディングの質問。
output string LLM の応答。
domain string "generic" または "algorithmic" のいずれか。
generation_algorithm string "self-instruct" または "evol-instruct" のいずれか。
llm_judgement string LLM による応答の評価を複数の基準で表した、JSON 辞書の文字列表現。
unit_tests string assert 文のリストの文字列表現。
tests_execution_status string "pass" または "fail" を示す文字列のリストの文字列表現。
average_test_score float 合格したテストケースの割合。

使い方

次の 2 行のコードでデータセットを読み込める。

from datasets import load_dataset
opencodeinstruct = load_dataset("nvidia/OpenCodeInstruct", split="train")

データセットの定量

  • レコード数 - 500 万件のコードの質問と回答のペア。
  • ダウンロードサイズ - 6.4 GB

倫理的な配慮:

NVIDIA は、信頼できる AI は共有の責任であると考えており、幅広い AI の応用を可能にするための方針と実践を定めている。利用規約に従ってダウンロード・利用する場合、開発者は、このモデルが関連する業界と用途の要件を満たし、想定外の製品の悪用に対応できるよう、社内のモデルチームと協力すべきである。

セキュリティ上の脆弱性や NVIDIA の AI に関する懸念は こちら から報告すること。

引用

このデータが役に立った場合は、次を引用してほしい:

@article{ahmad2025opencodeinstruct,
      title={OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs}, 
      author={Wasi Uddin Ahmad and Aleksander Ficek and Mehrzad Samadi and Jocelyn Huang and Vahid Noroozi and Somshubra Majumdar and Boris Ginsburg},
      year={2025},
      eprint={2504.04030},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.04030}, 
}
原文(英語)

OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs

Dataset Description

We introduce OpenCodeInstruct, the largest open-access instruction tuning dataset, comprising 5 million diverse samples. OpenCodeInstruct is designed for supervised fine-tuning (SFT).

  • Technical Report - Discover the methodology and technical details behind OpenCodeInstruct.
  • Github Repo - Access the complete pipeline used to perform SFT.

This dataset is ready for commercial/non-commercial use.

Dataset Owner(s)

NVIDIA Corporation

Dataset Creation Date

January 2025 - March 2025

License/Terms of Use

GOVERNING TERMS: This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0) available at https://creativecommons.org/licenses/by/4.0/legalcode.

Data Developer: NVIDIA

Use Case:

Developers training LLMs to specialize LLMs in code generation.

Release Date:

04/28/2025

Intended Usage

The OpenCodeInstruct Dataset is intended to be used by the community to continue to improve open models. The data may be freely used to train models. However, for each dataset a user elects to use, the user is responsible for checking if the dataset license is fit for the intended purpose.

Dataset Characterization

** Data Collection Method

  • [Hybrid: Automated, Synthetic]

** Labeling Method

  • [Hybrid: Automated, Synthetic]

Dataset Format

Field Type Description
id string A unique id for each question
input string The input coding question.
output string LLM's response.
domain string Either "generic" or "algorithmic".
generation_algorithm string Either "self-instruct" or "evol-instruct".
llm_judgement string string representation of a JSON dictionary containing the LLM's evaluation of the response based on several criteria.
unit_tests string string representation of a list of assertion statements.
tests_execution_status string string representation of a list of strings indicating "pass" or "fail".
average_test_score float Fraction of test cases passed.

How to Use It

You can load the dataset with the following two lines of code.

from datasets import load_dataset
opencodeinstruct = load_dataset("nvidia/OpenCodeInstruct", split="train")

Dataset Quantification

  • Record Count - 5 million coding question-answer pairs.
  • Download Size - 6.4 GB

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Citation

If you find the data useful, please cite:

@article{ahmad2025opencodeinstruct,
      title={OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs}, 
      author={Wasi Uddin Ahmad and Aleksander Ficek and Mehrzad Samadi and Jocelyn Huang and Vahid Noroozi and Somshubra Majumdar and Boris Ginsburg},
      year={2025},
      eprint={2504.04030},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.04030}, 
}