IntelliProcure/SwissSPARK

⚠️ Caution: The dataset is subject to continuous changes. We are currently actively developing it.

Dataset Card: Dataset for a Swiss Sustainable Procurement Analysis & Reporting Kit

Dataset Description

This dataset is designed to train and evaluate models for detecting sustainability criteria in Swiss public procurement documents (Call for Tenders, CFT). The dataset classifies text segments based on whether they contain specific sustainability requirements according to official Swiss guidelines.

These annotations are based on a specific snapshot of the sustainability procurement criteria catalogs available at IntelliProcure/sustainability_criteria. However, to ensure consistency and reproducibility with the SwissSPARK annotations, users are advised to use this specific snapshot of the criteria catalogs: IntelliProcure/SwissSPARK_Catalogs. Newer versions of the catalogs may contain changes that differ from the guidelines used during the annotation of this dataset.

Dataset Summary

• Language: German (de)
• Task: Multi-label classification (sustainability criteria detection, procurement criterion classification, sector-specific classification)
• Domain: Public procurement, sustainability criteria
• Size: Multiple splits with stratified sampling

Supported Tasks

• Sustainability Criteria Detection: Identify whether text segments contain sustainability requirements (via area_of_action_id mapping)
• Sector-Specific Classification: Classify texts according to sector-specific guidelines (catalog)
• Procurement Criterion Classification: Determine the legal context of criteria (e.g., technical_specification, award_criterion, selection_criterion, terms_and_conditions, evidence_requirement)

Dataset Structure

Data Fields

• project_id (int): Unique identifier for the procurement project
• simap_version (string): Version of SIMAP platform (e.g., "simap")
• filename (string): Original document filename
• source (string): Data source ("annotations" or "feedback")
• text (string): The original text segment
• text_cleaned (string): Normalized and cleaned text
• catalog (string): Sector-specific guideline (e.g., "Road_Transport", "Print_Services", "ICT", "Food", etc.)
• area_of_action_id (string): Specific criterion ID within the catalog (e.g., "HF-001", or "not present" for negative examples)
• sustainability_dimensions (string): Sustainability dimensions of the annotated area of action as a bracketed list, e.g. "[ecologic, social]" ("[]" for negative examples or when no dimension applies)
• procurement_criterion (string): Legal context type (e.g., "technical_specification", "award_criterion", "selection_criterion", "terms_and_conditions", "evidence_requirement", or "not present")
• self_declaration (string): Indicates if it's a self-declaration ("True"/"False" as string, or "not present")
• evidence (string): Evidence requirements (or "not present")
• award_criterion_weight (string): Weight information for award criteria (or "not present")

Data Splits

The dataset provides a coverage-aware, leakage-safe split:

• 50% Training
• 25% Evaluation
• 25% Test

Splitting happens at the document-snippet level: a snippet and all of its annotation rows always land in the same split. Two seeded (and therefore reproducible) phases then allocate the snippet groups:

1. Coverage — every (catalog, area_of_action_id) pair of the positively annotated snippets and every real procurement_criterion value is processed rarest-first and placed into splits that do not contain it yet. Pairs held by 3 or more snippets are guaranteed to appear in all three splits; rarer pairs cover as many splits as possible and are flagged in the Split_Coverage sheet of statistics.xlsx.
2. Balancing — the remaining snippets go to the split where their labels lag furthest behind the ideal 50/25/25 quota, so procurement_criterion shares typically deviate by less than 1 % from the ideal proportions.

statistics.xlsx additionally reports how the annotations are distributed across the splits at row level: the Split_Areas_Of_Action, Split_Catalogs and Split_Criteria sheets each list every label with its total annotation count, its count per split and the share of the label's annotations that each split receives.

The sentinels "not present"/"not specified" never count as labels; negative snippets carry no labels and are distributed purely by remaining split capacity.

Sector-Specific Guidelines (Catalogs)

The dataset includes sustainability criteria from multiple sectors:

• Road Transport
• Food
• ICT
• Furniture
• Construction (Buildings/Civil Engineering)
• Electricity
• Print Services
• Product Independent Criteria
• And others

Metadata Subset (project_summaries)

Besides the annotated text data, the repository publishes a per-project metadata subset — project_summaries.json / project_summaries.parquet, one row per procurement project (~56 rows). It documents who annotated what and which sector catalogs came into question, so the text data can be put into project context without exposing the annotation workflow's internal bookkeeping. On Hugging Face the subset is published as its own config/split pair — config metadata, split metadata (see the configs: block above) — so the Dataset Viewer shows it in the split selector next to the default config's train/eval/test splits, and load_dataset("IntelliProcure/SwissSPARK", "metadata", split="metadata") loads it directly:

Field	Type	Description
project_id	int	SIMAP project ID
simap_version	string	SIMAP platform version (e.g. "simap")
division / cpv_codes	string	CPV division text / CPV codes of the project
annotators	list of strings	Pseudonymized annotator IDs (see below)
candidate_catalogs	list of strings	Catalogs whose workbook column is marked match — they came into question for this project
annotated_catalogs	list of strings	Catalogs marked annotated — they were actually annotated for this project
documents	list of strings	The project's documents (dataset filenames, joinable to the main subset's filename)

Catalog names equal the catalog vocabulary of the default config (e.g. Food, Road_Transport, Product_Independent_Criteria).

Annotator pseudonymization: annotator names are never published raw. For each release, all annotators recorded in the source workbook are sorted alphabetically and mapped onto annotator_1 … annotator_N; the mapping is identical in every row. The IDs are release-scoped: they are regenerated on every rebuild and the numbering can shift between releases when the annotator roster changes. Columns that may contain personal data ("Recommended by", internal comments) are never exported, and generation fails hard if a real annotator name is detected anywhere in the subset.

Dataset Creation

Source Data

The dataset combines two primary sources:

1. Expert Annotations (INCEpTION Platform): CFT documents manually labeled by experts on the UIMA INCEpTION platform with precise mapping to catalog IDs (area_of_action_id) and criterion types (procurement_criterion)

2. Feedback Data: 220 examples validated by experts from an automated NLP pipeline, including weight information for award criteria (award_criterion_weight) and correctness evaluations

Data Processing Pipeline

The generation pipeline includes:

1. Positive Example Extraction: Direct extraction from expert-annotated documents
2. Negative Sampling:
   • Hard Negatives: Sentences with high semantic similarity (cosine similarity) but no sustainability criteria
   • Random Negatives: Random sentences for standard administrative language
3. Contextual Windowing: Creates context windows (1 or 2 surrounding sentences) around each excerpt; negative candidates overlapping annotated excerpts are skipped during mining instead of being published with an in-context warning
4. Text Cleaning: Normalization of line breaks, removal of excess whitespace (available in text_cleaned)
5. Deduplication: Based on project_id, simap_version, text_cleaned, filename, catalog, and area_of_action_id

Annotations

Annotations were created by domain experts familiar with:

• Swiss public procurement law
• Sustainability guidelines for public procurement
• Sector-specific requirements

Annotation categories (procurement_criterion) include:

• award_criterion (Zuschlagskriterien)
• selection_criterion (Eignungskriterien)
• technical_specification (Technische Spezifikationen)
• terms_and_conditions (Vertragsbedingungen)
• evidence_requirement (Nachweise)

Technical Specifications

• NLP Engine: spaCy (model: de_dep_news_trf) for German sentence segmentation
• Embeddings: SentenceTransformers (all-MiniLM-L6-v2) for semantic search and hard negative mining
• Language filter: lingua (restricted to German/French/Italian/English) excludes non-German sentences from mined negative examples
• Environment: Python 3.10 via condavenv

Considerations

Social Impact

This dataset supports:

• Transparency: Automated detection of sustainability criteria in public procurement
• Efficiency: Reduces manual review time for procurement officers
• Sustainability: Promotes adherence to sustainability guidelines in Swiss public procurement

Limitations

• Language: Dataset is German-only, specific to Swiss German administrative language (mined negative examples are filtered to German-only sentences via lingua; human-annotated positives are kept as annotated)
• Domain Specificity: Trained on Swiss public procurement documents; may not generalize to other jurisdictions
• Context Dependency: Some criteria require understanding of broader document context
• Class Imbalance: Distribution between positive and negative examples should be monitored during training (mapped via area_of_action_id != "not present")
• Release-Scoped Annotator Pseudonyms: the annotator_* IDs of the project_summaries config are recomputed from the source workbook on every release and may change between releases; they identify a person only within one published version

Citation

If you use this dataset, please cite:

@dataset{swiss_sustainable_procurement_2026,
  author       = { Veton Matoshi and Rahel Meili and Judith Binder and Sarah Hostettler and Tilia Ellendorff },
  title        = { SwissSPARK },
  year         = 2026,
  url          = { https://huggingface.co/datasets/IntelliProcure/SwissSPARK },
  doi          = { 10.57967/hf/7945 },
  publisher    = { Hugging Face }
}

License

[Please specify license information based on your organization's requirements]

Contact

For questions or feedback regarding this dataset, please contact [contact information].