openeurollm/ground-truth-bsc-edu-annotations

Dataset Card for Ground Truth BSC-EDU Dataset

The Ground Truth BSC-EDU Dataset is a multilingual collection of 3,000 manually annotated webpages for educational quality assessment. All records include primary annotations from the main project, with additional alternative annotations from re-annotation and validation projects for selected documents. Annotations follow a 0-4 quality scale and include quality flags for problematic content. This dataset serves as the gold standard for training and evaluating LLM-based quality scoring systems.

Dataset Details

Dataset Description

This dataset contains 3,000 human-annotated webpages from Common Crawl across three languages: Spanish, Catalan, and Basque. All records include primary annotations from the main project (project_1000), plus optional alternative annotations from re-annotation projects for subset of documents.

Each webpage is scored on a 0-4 scale reflecting educational value and includes binary quality flags. All records include the primary annotation fields (label, label_int, flags), with additional columns for alternative annotations (label-first200, label-rest50, label-external10) containing data only for documents that were re-annotated. This structure enables straightforward training on primary annotations while supporting detailed inter-annotator agreement analysis.

The dataset is organized into subsets by language (Spanish, Catalan, Basque) with an "all" subset combining all languages. Train/test splits (90%/10%) are provided for each language based on the primary annotations, with additional annotation data available for thorough evaluation and model calibration.

• Curated by: AI Institute, Barcelona Supercomputing Center
• Funded by: OpenEuroLLM consortium (Grant Agreement No 101195233) and Ministerio para la Transformación Digital y de la Función Pública (ALIA)
• Language(s) (NLP): Spanish (spa_Latn), Catalan (cat_Latn), Basque (eus_Latn)
• License: CC0 1.0 (annotations and metadata only)

Dataset Sources

• Repository: https://huggingface.co/datasets/openeurollm/ground-truth-bsc-edu-dataset
• Paper: -
• Source Data: FineWeb-2 collection (Common Crawl)

Uses

Direct Use

This dataset is suitable for:

• Training quality scoring classifiers for educational content assessment
• Validating and benchmarking LLM-based annotation systems
• Fine-tuning regression models for continuous quality scoring
• Developing and evaluating multilingual quality filtering pipelines
• Testing model generalization across languages (Spanish, Catalan, Basque)

Out-of-Scope Use

• Using annotations to make judgments beyond educational quality
• Reproducing or redistributing copyrighted webpage content
• Making claims about webpage accuracy, relevance, or other quality dimensions
• Content moderation or hate speech detection (outside annotation scope)

Dataset Structure

Data Fields

Each JSONL record contains:

Metadata:

• ref_id (string): Unique reference identifier
• url (string): Source URL from Common Crawl
• file_path (string): Common Crawl S3 file path
• lang (string): Language code (cat, es, eu)
• lang_script (string): ISO 15924 script code (Latn for Latin)
• lang_score (float): Language detection confidence (0-1)
• top_langs (string): JSON-formatted language detection scores
• minhash_cluster_size (integer): Deduplication cluster size

Annotations (Primary - from project_1000):

• label (string): Quality label name (None, Minimal, Basic, Good, Excellent)
• label_int (integer): Quality score (0-4)
• flags (array): List of applicable quality flags
• label_created_at (string): ISO 8601 timestamp of annotation creation
• label_updated_at (string): ISO 8601 timestamp of annotation update

Annotations (Optional Re-annotations - null for documents not re-annotated):

• label-first200 (object or null): Re-annotation from project-200 (600 documents, 20%), with value, score, flags subfields
• label-rest50 (object or null): Re-annotation from project-50 (150 documents, 5%), with value, score, flags subfields
• label-external10 (object or null): External validation from project-10 (30 documents, 1%), with value, score, flags subfields

Data Splits

The dataset provides train/test splits (90%/10% ratio) for each language based on the main annotation project (project_1000). All records include columns for alternative annotations from re-annotation and validation projects, with null values for documents that were not re-annotated.

Train/Test Splits (Based on Main Annotation Project):

Subset	Train	Test	Total
Spanish (spa_Latn)	900	100	1,000
Catalan (cat_Latn)	900	100	1,000
Basque (eus_Latn)	900	100	1,000
All Combined	2,700	300	3,000

Multi-Annotation Coverage:

All 3,000 records contain primary annotations (label, label_int, flags from project_1000). Re-annotation columns include data only for documents that were annotated by those projects:

Re-annotation Project	Annotated Documents	Coverage
label-first200	600 documents	20%
label-rest50	150 documents	5%
label-external10	30 documents	1%

For documents without re-annotations, the corresponding columns contain null values. A total of 767 documents (26% of dataset) have annotations from multiple projects, enabling inter-annotator agreement analysis.

Dataset Creation

Curation Rationale

As LLM-based content filtering becomes standard practice in LLM pretraining, the validity of LLM annotators for subjective quality criteria remains largely unexplored, especially for non-English languages. This dataset was created to:

1. Establish a reliable multilingual gold standard for educational quality assessment
2. Enable systematic evaluation of LLM annotators against human judgments
3. Support training of efficient, specialized quality scoring models
4. Facilitate reproducible research in multilingual content filtering

Source Data

Data Collection and Processing

Webpages were selected from Common Crawl and preprocessed as follows:

1. Language identification filtering for Spanish, Catalan, and Basque
2. Text extraction and normalization
3. Balanced sampling to ensure language distribution
4. JSONL format organization for streaming efficiency

Who are the source data producers?

Webpages originate from Common Crawl (https://commoncrawl.org/), a non-profit organization that regularly crawls the public web and makes the data openly available.

Annotations

Annotation process

The dataset was created through multiple annotation phases with documented project-based annotations:

Phase 1 (main project): Initial annotation of base 3,000 webpages

• Annotators scored each webpage on the 0-4 educational quality scale
• Marked applicable quality flags (adult content, incorrect language, problematic content)
• Quality assurance through peer review
• Produces train/test splits used for standard model evaluation

Phase 2 (label-first200 project): Re-annotation of first 200 instances per language

• Independent annotators re-annotated 200 instances from each language
• Purpose: Validate annotation quality and measure inter-annotator agreement

Phase 3 (label-rest50 & label-external10 projects): Additional validation

• Re-annotation of next 50 instances per language (label-rest50)
• External expert review of 10 instances per language (label-external10)

Annotation Guidelines:

Educational Quality Scale (0-4):

• 0 - None/Minimal: Promotional, personal, or entertainment content with no educational intent
• 1 - Minimal: Topics mentioned or highlighted without explanation or depth
• 2 - Basic: Basic facts or definitions provided without conceptual development
• 3 - Good: Concepts clearly explained with relationships, processes, or examples
• 4 - Excellent: Detailed explanations with instructions, examples, or practical exercises

Quality Flags (binary):

• Adult content: Inappropriate or sexual content
• Incorrect language: Text in wrong language or machine translation artifacts
• Problematic content: Illegible, corrupted, automatically generated, or error messages

Who are the annotators?

Annotations were created by trained human annotators with domain expertise in education and language. All annotators were trained on the annotation guidelines and achieved agreement thresholds before annotation began. Re-annotation projects involved independent annotators to measure inter-annotator agreement.

Personal and Sensitive Information

The dataset contains webpages from Common Crawl that may include:

• Publicly accessible personal information (names, email addresses)
• URLs and web references
• Potentially sensitive information in webpage content

Users should:

• Comply with Common Crawl's Terms of Use
• Respect privacy regulations in their jurisdiction
• Obtain necessary permissions before reproducing content
• Consider ethical implications of using public web data

Bias, Risks, and Limitations

Technical Limitations

• Language coverage: Dataset limited to three Iberian languages (Spanish, Catalan, Basque)
• Educational focus: Scoring measures educational value only; does not assess accuracy, relevance, or other quality dimensions
• Common Crawl bias: Reflects biases inherent in web crawling and content availability
• Annotation disagreement: Some webpages may have unclear educational quality; inter-annotator agreement varies by category

Recommended Mitigations

• Report agreement metrics when using annotations for model evaluation
• Use train/test splits to avoid overfitting and assess generalization
• Consider language-specific characteristics when extending to other languages
• Combine with other quality signals for robust filtering pipelines
• Document model predictions and disagreement cases for transparency

Citation

BibTeX:

More Information

Detailed Annotation Guidelines: See GUIDELINES.md for comprehensive annotation criteria and examples.

Frequently Asked Questions: See Q&A.xlsx for common questions about the annotation process.

Related Datasets:

• LLM-Judge EDU Dataset: 1.5 million webpages annotated by LLM judges
• Annotated HPLT 4.0: 38 languages (~11T tokens) filtered using quality classifiers trained on this dataset

Dataset Card Contact

Barcelona Supercomputing Center (BSC), Spain

Acknowledgments

Co-funded by the European Union (Grant Agreement No 101195233). Views and opinions expressed are those of the OpenEuroLLM consortium only and do not necessarily reflect those of the European Union or Digital Europe Programme. Neither the European Union nor the granting authority can be held responsible for them.

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA.