Files changed (1) hide show
  1. README.md +17 -17
README.md CHANGED
@@ -19,13 +19,13 @@ task_categories:
19
  ---
20
  # Nemotron-PII-Japan Dataset Card
21
 
22
- 日本語PII/PHI検出・編集のための合成データ / Synthetic Japanese Data for PII/PHI Detection and Redaction
23
 
24
  # データセット概要 (Dataset Overview)
25
 
26
- Nemotron-PII-Japanは、日本語テキスト内の個人識別情報(PII)および保護対象となり得る健康・医療情報(PHI)の検出、抽出、編集、評価を目的とした、ペルソナに基づく合成データセットです。日本の氏名、住所、識別子、金融慣習、組織、業務文書などの文脈を扱い、実在する機密性の高い文書に依存しないモデル開発を支援します。本カードにおけるPHIは記述上の分類であり、HIPAAまたは日本の法令への適合を意味しません。
27
 
28
- Nemotron-PII-Japan is a synthetic, persona-grounded dataset for training and evaluating the detection, extraction, and redaction of personally identifiable information (PII) and potentially protected health information (PHI) in Japanese text. It covers Japanese names, addresses, identifiers, financial conventions, organizations, and business-document contexts without relying on real-world sensitive or confidential documents. In this card, PHI is used as a descriptive category and does not imply compliance with HIPAA or Japanese law.
29
 
30
  本データセットは [NVIDIA NeMo Data Designer](https://github.com/NVIDIA-NeMo/DataDesigner) を用いて生成されます。人物属性には [Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) を基盤とするData Designer向け日本語persona assetを使用し、固定形式の値は決定的なsampler・expression・PIIプラグイン、地域固有の名称は管理された合成カタログ、自然な文書本文はLLM生成と検証・正規化を組み合わせて作成します。
31
 
@@ -55,7 +55,7 @@ This dataset is ready for commercial or non-commercial uses.
55
 
56
  # ユースケース (Use Cases)
57
 
58
- * 日本語PII/PHIの固有表現認識と抽出 / Japanese PII/PHI named-entity recognition and extraction
59
  * PIIの黒塗り、匿名化、編集処理の学習・評価 / Training and evaluation of PII redaction, anonymization, and editing
60
  * Data Loss Prevention、検索、RAG、エージェント、会話AIの安全性評価 / Safety evaluation for DLP, search, RAG, agents, and conversational AI
61
  * 構造化・非構造化の日本語文書に対する文書分類と堅牢性評価 / Document classification and robustness testing across structured and unstructured Japanese text
@@ -155,33 +155,33 @@ Each exported record contains the following fields:
155
  8つの業務分野に属する43種類の日本語文書を対象としています。
156
  対象分野は、医療、金融、行政、雇用、教育、法務、通信・IT、およびECです。
157
  氏名、住所、連絡先、マイナンバー、金融・医療・デジタル識別子、日付・日時・時刻など、56種類のPIIラベル(医療機関名・教育機関名・公的機関名を含む)に対応し、810,703スパンを含みます。
158
- 各レコードには、1件の文書と複数の文字単位PIIアノテーションが含まれます。data.parquetは75.52 MB(10進MB)です。
159
 
160
  It covers 43 Japanese document types across eight business domains.
161
  The covered domains are Healthcare, Finance, Government, Employment, Education, Legal, Telecommunications and IT, and E-commerce.
162
  The dataset supports 56 native PII labels (including medical and educational institution names and public-body names), including names, addresses, contact information, My Number, financial, medical and digital identifiers, dates, datetimes and times, with 810,703 annotated spans.
163
- Each record contains one document and multiple character-level PII annotations. The data.parquet file is 75.52 MB (decimal MB).
164
 
165
  # データセット詳細 (Dataset Details)
166
 
167
  ## オープンデータとしての特徴 / Open-data characteristics
168
 
169
  * 言語: 日本語(ja) / Language: Japanese (ja)
170
- * タスク: token classification、PII/PHI抽出、redaction evaluation / Tasks: token classification, PII/PHI extraction, and redaction evaluation
171
  * 生成・注釈方法: persona sampler、決定的sampler・expression、PIIプラグイン、合成カタログなどでPIIの値を先に生成し、その値を含む日本語文書をLLMで生成します。確定した本文に対してルールベースの後処理を行い、文字単位の範囲(span)とラベルを付与します。/ Generation and annotation: PII values are generated first using person samplers, deterministic samplers and expressions, PII plug-ins, and synthetic catalogs. An LLM then generates Japanese documents incorporating those values. Rule-based post-processing assigns character-level spans and labels to the finalized text.
172
- * 人物範囲: 成人のみ(JP profileでは18~74歳) / Population scope: adults only (ages 18–74 in the JP profile)
173
  * ドメイン: 8 / Domains: 8
174
  * 総件数: 100,000件 / Total count: 100,000 records /
175
  * 学習taxonomyの文書タイプ: 43 / Document types in the training taxonomy: 43
176
  * PIIラベル: 56 / Native PII labels: 56
177
  * レコード数: 100,000 / Record count: 100,000
178
  * 保存形式: data.parquet、100,000文書の単一ファイル / Storage layout: data.parquet, a single file of 100,000 documents
179
- * ストレージサイズ: data.parquet 75.52 MB(10進MB) / Storage: data.parquet, 75.52 MB (decimal MB)
180
- * トークン数: 34,737,684 / Token count: 34,737,684
181
  * トークン数は、Nemotron Nano v2 Japaneseのトークナイザーを使用し、特殊トークンとチャットテンプレートを除外して、文書本文から直接計測しました / The token counts were measured directly from the document text using the tokenizer for Nemotron Nano v2 Japanese, excluding special tokens and chat templates.
182
  * PIIスパン数:810,703(約81.1万)。本文中のPIIに付けた位置・ラベル注釈の総数で、1文書平均約8.11箇所です。/ PII span count: 810,703 (approximately 811,000). This is the total number of position and label annotations attached to PII in the document text、 averaging approximately 8.11 spans per document.
183
  * ラベルの範囲: 注釈は、生成スロットとして用意したPII値と、ルールで検出した範囲を中心としています。本文中の同種の情報をすべて網羅するものではなく、スロット以外の値やルールの対象外となる表現には、ラベルが付いていない場合があります。/ Annotation coverage: Labels primarily cover PII values prepared in generation slots and spans detected by rules. Coverage is not exhaustive: other values of the same type, or expressions outside the rules, may remain unannotated.
184
- * 文脈による注釈範囲: 機関・施設の所在地、機器・文書のID、銀行などの組織情報には、注釈方針や出現箇所の文脈に応じて、ラベルが付く場合と付かない場合があります。これらを一律に対象外としているわけではありません。この区別はPIIの法的な定義ではなく、本データセットの注釈範囲・制約を示します。/ Context-dependent scope: Institutional or facility locations, equipment or document IDs, and organizational information such as bank names may be labeled or left unannotated depending on the annotation policy and occurrence context. These categories are not categorically excluded. This describes the dataset's annotation scope and limitations, rather than a legal definition of PII.
185
 
186
  ## 対象ドメイン (Domains)
187
 
@@ -213,7 +213,7 @@ The following are three synthetic samples covering medical, finance and public-a
213
  ## 医療 (Medical)
214
 
215
  文書タイプ / Document type: 診療情報提供書 (Medical Referral Letter)
216
- V12\_2収録サンプル / Sample from V12\_2 (UID: b9d76902728a43699eb182b07d298f3b):
217
 
218
  ```
219
  診療情報提供書
@@ -249,7 +249,7 @@ V12\_2収録サンプル / Sample from V12\_2 (UID: b9d76902728a43699eb182b07d29
249
  ## 金融 (Finance)
250
 
251
  文書タイプ / Document type: 口座開設申込書 (Account Opening Application)
252
- V12\_2収録サンプル / Sample from V12\_2 (UID: 6ee53c8431ae4100a63335757cd89c5d):
253
 
254
  ```
255
  口座開設申込書
@@ -280,7 +280,7 @@ V12\_2収録サンプル / Sample from V12\_2 (UID: 6ee53c8431ae4100a63335757cd8
280
  ## 行政 (Public Administration)
281
 
282
  文書タイプ / Document type: 問い合わせ受付票 (Inquiry Intake Form)
283
- V12\_2収録サンプル / Sample from V12\_2 (UID: 071cc5d287fe4af4a8fcc9b121f672a2):
284
 
285
  ```
286
  問い合わせ受付票
@@ -318,9 +318,9 @@ V12\_2収録サンプル / Sample from V12\_2 (UID: 071cc5d287fe4af4a8fcc9b121f6
318
 
319
  ## 人物データ (Persona Data)
320
 
321
- JP Profileは、[Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan)を基盤とする100万件のData Designer向け日本語person assetを、標準のPerson Sampler経由で使用します。氏名、年齢、生年月日、性別、職業、電話、メール、都道府県、市区町村、郵便番号などを、同一人物内で整合する生成基盤として利用します。
322
 
323
- JP Profile uses a 1,000,000-record Japanese Data Designer persona asset grounded in [Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) through the standard Person Sampler. It provides names, ages, dates of birth, sex, occupations, phone numbers, email addresses, prefectures, cities, postcodes, and other attributes as a consistent per-person generation foundation.
324
 
325
  # シードデータ (Seed Data)
326
 
@@ -411,7 +411,7 @@ Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns \[[h
411
  ```
412
  @dataset{nemotron-pii-japan,
413
  author = {Ogushi, Masaya and Fujita, Atsunori and Gong, Vincent and Yamamoto, Kotaro and Ruan, Colleen and Suhara, Yoshi and Jennings, Will and Manoel, Andre},
414
- title = {Nemotron-PII-Japan: Synthetic Japanese Data for PII/PHI Detection and Redaction},
415
  year = {2026},
416
  publisher = {NVIDIA},
417
  url = {https://huggingface.co/datasets/nvidia/Nemotron-PII-Japan}
 
19
  ---
20
  # Nemotron-PII-Japan Dataset Card
21
 
22
+ 日本語PII・編集のための合成データ / Synthetic Japanese Data for PII Detection and Redaction
23
 
24
  # データセット概要 (Dataset Overview)
25
 
26
+ Nemotron-PII-Japanは、日本語テキスト内の個人識別情報(PII)の検出、抽出、編集、評価を目的とした、ペルソナに基づく合成データセットです。日本の氏名、住所、識別子、金融慣習、組織、業務文書などの文脈を扱い、実在する機密性の高い文書に依存しないモデル開発を支援します。
27
 
28
+ Nemotron-PII-Japan is a synthetic, persona-grounded dataset for training and evaluating the detection, extraction, and redaction of personally identifiable information (PII) in Japanese text. It covers Japanese names, addresses, identifiers, financial conventions, organizations, and business-document contexts without relying on real-world sensitive or confidential documents.
29
 
30
  本データセットは [NVIDIA NeMo Data Designer](https://github.com/NVIDIA-NeMo/DataDesigner) を用いて生成されます。人物属性には [Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) を基盤とするData Designer向け日本語persona assetを使用し、固定形式の値は決定的なsampler・expression・PIIプラグイン、地域固有の名称は管理された合成カタログ、自然な文書本文はLLM生成と検証・正規化を組み合わせて作成します。
31
 
 
55
 
56
  # ユースケース (Use Cases)
57
 
58
+ * 日本語PIIの固有表現認識と抽出 / Japanese PII named-entity recognition and extraction
59
  * PIIの黒塗り、匿名化、編集処理の学習・評価 / Training and evaluation of PII redaction, anonymization, and editing
60
  * Data Loss Prevention、検索、RAG、エージェント、会話AIの安全性評価 / Safety evaluation for DLP, search, RAG, agents, and conversational AI
61
  * 構造化・非構造化の日本語文書に対する文書分類と堅牢性評価 / Document classification and robustness testing across structured and unstructured Japanese text
 
155
  8つの業務分野に属する43種類の日本語文書を対象としています。
156
  対象分野は、医療、金融、行政、雇用、教育、法務、通信・IT、およびECです。
157
  氏名、住所、連絡先、マイナンバー、金融・医療・デジタル識別子、日付・日時・時刻など、56種類のPIIラベル(医療機関名・教育機関名・公的機関名を含む)に対応し、810,703スパンを含みます。
158
+ 各レコードには、1件の文書と複数の文字単位PIIアノテーションが含まれます。data.parquetは75.52 MBです。
159
 
160
  It covers 43 Japanese document types across eight business domains.
161
  The covered domains are Healthcare, Finance, Government, Employment, Education, Legal, Telecommunications and IT, and E-commerce.
162
  The dataset supports 56 native PII labels (including medical and educational institution names and public-body names), including names, addresses, contact information, My Number, financial, medical and digital identifiers, dates, datetimes and times, with 810,703 annotated spans.
163
+ Each record contains one document and multiple character-level PII annotations. The data.parquet file is 75.52 MB.
164
 
165
  # データセット詳細 (Dataset Details)
166
 
167
  ## オープンデータとしての特徴 / Open-data characteristics
168
 
169
  * 言語: 日本語(ja) / Language: Japanese (ja)
170
+ * タスク: token classification、PII抽出、redaction evaluation / Tasks: token classification, PII extraction, and redaction evaluation
171
  * 生成・注釈方法: persona sampler、決定的sampler・expression、PIIプラグイン、合成カタログなどでPIIの値を先に生成し、その値を含む日本語文書をLLMで生成します。確定した本文に対してルールベースの後処理を行い、文字単位の範囲(span)とラベルを付与します。/ Generation and annotation: PII values are generated first using person samplers, deterministic samplers and expressions, PII plug-ins, and synthetic catalogs. An LLM then generates Japanese documents incorporating those values. Rule-based post-processing assigns character-level spans and labels to the finalized text.
172
+ * 人物範囲: 成人のみ(18~74歳) / Population scope: adults only (ages 18–74)
173
  * ドメイン: 8 / Domains: 8
174
  * 総件数: 100,000件 / Total count: 100,000 records /
175
  * 学習taxonomyの文書タイプ: 43 / Document types in the training taxonomy: 43
176
  * PIIラベル: 56 / Native PII labels: 56
177
  * レコード数: 100,000 / Record count: 100,000
178
  * 保存形式: data.parquet、100,000文書の単一ファイル / Storage layout: data.parquet, a single file of 100,000 documents
179
+ * ストレージサイズ: data.parquet 75.52 MB / Storage: data.parquet, 75.52 MB
180
+ * トークン数: 34,737,684 (約3,474万)/ Token count: 34,737,684 (approximately 34,740,000)
181
  * トークン数は、Nemotron Nano v2 Japaneseのトークナイザーを使用し、特殊トークンとチャットテンプレートを除外して、文書本文から直接計測しました / The token counts were measured directly from the document text using the tokenizer for Nemotron Nano v2 Japanese, excluding special tokens and chat templates.
182
  * PIIスパン数:810,703(約81.1万)。本文中のPIIに付けた位置・ラベル注釈の総数で、1文書平均約8.11箇所です。/ PII span count: 810,703 (approximately 811,000). This is the total number of position and label annotations attached to PII in the document text、 averaging approximately 8.11 spans per document.
183
  * ラベルの範囲: 注釈は、生成スロットとして用意したPII値と、ルールで検出した範囲を中心としています。本文中の同種の情報をすべて網羅するものではなく、スロット以外の値やルールの対象外となる表現には、ラベルが付いていない場合があります。/ Annotation coverage: Labels primarily cover PII values prepared in generation slots and spans detected by rules. Coverage is not exhaustive: other values of the same type, or expressions outside the rules, may remain unannotated.
184
+ * 文脈による注釈範囲: 機関・施設の所在地、機器・文書のID、銀行などの組織情報には、注釈方針や出現箇所の文脈に応じて、ラベルが付く場合と付かない場合があります。これらを一律に対象外としているわけではありません。この区別はPIIの法的な定義ではなく、本データセットの注釈範囲・制約を示します。また、注釈対象に該当する情報でも、ラベルの欠落や範囲・分類の誤りが残っている可能性があります。/ Context-dependent scope: Institutional or facility locations, equipment or document IDs, and organizational information such as bank names may be labeled or left unannotated depending on the annotation policy and occurrence context. These categories are not categorically excluded. This describes the dataset's annotation scope and limitations, rather than a legal definition of PII. Additionally, information within the annotation scope may still have missing labels, incorrect span boundaries, or misclassified labels.
185
 
186
  ## 対象ドメイン (Domains)
187
 
 
213
  ## 医療 (Medical)
214
 
215
  文書タイプ / Document type: 診療情報提供書 (Medical Referral Letter)
216
+ 収録サンプル / Sample from (UID: b9d76902728a43699eb182b07d298f3b):
217
 
218
  ```
219
  診療情報提供書
 
249
  ## 金融 (Finance)
250
 
251
  文書タイプ / Document type: 口座開設申込書 (Account Opening Application)
252
+ 収録サンプル / Sample from (UID: 6ee53c8431ae4100a63335757cd89c5d):
253
 
254
  ```
255
  口座開設申込書
 
280
  ## 行政 (Public Administration)
281
 
282
  文書タイプ / Document type: 問い合わせ受付票 (Inquiry Intake Form)
283
+ 収録サンプル / Sample from (UID: 071cc5d287fe4af4a8fcc9b121f672a2):
284
 
285
  ```
286
  問い合わせ受付票
 
318
 
319
  ## 人物データ (Persona Data)
320
 
321
+ Persona Dataは、[Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan)を基盤とする100万件のData Designer向け日本語person assetを、標準のPerson Sampler経由で使用します。氏名、年齢、生年月日、性別、職業、電話、メール、都道府県、市区町村、郵便番号などを、同一人物内で整合する生成基盤として利用します。
322
 
323
+ Persona Data uses a 1,000,000-record Japanese Data Designer persona asset grounded in [Nemotron-Personas-Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) through the standard Person Sampler. It provides names, ages, dates of birth, sex, occupations, phone numbers, email addresses, prefectures, cities, postcodes, and other attributes as a consistent per-person generation foundation.
324
 
325
  # シードデータ (Seed Data)
326
 
 
411
  ```
412
  @dataset{nemotron-pii-japan,
413
  author = {Ogushi, Masaya and Fujita, Atsunori and Gong, Vincent and Yamamoto, Kotaro and Ruan, Colleen and Suhara, Yoshi and Jennings, Will and Manoel, Andre},
414
+ title = {Nemotron-PII-Japan: Synthetic Japanese Data for PII Detection and Redaction},
415
  year = {2026},
416
  publisher = {NVIDIA},
417
  url = {https://huggingface.co/datasets/nvidia/Nemotron-PII-Japan}