TENDRIL is released under the TENDRIL Open RAIL-D License (see LICENSE.md). Its use-based restrictions (Attachment A) include a ban on extracting personal data or re-identifying persons, and a duty to disclose training on the test collection. They must be passed on to anyone who receives the data or a derivative.

TENDRIL: Tender Retrieval Benchmark

TENDRIL is a German retrieval benchmark for finding sustainability-relevant passages in Swiss public procurement tenders. Each query describes an area of action (Handlungsfeld) from our sustainability catalogs, which are based on Swiss and EU recommendations for sustainable public procurement. Within one tender, the task is to retrieve the passages that address that area. The query text is a German summary of the criteria of the area of action, generated with DeepSeek V4 Flash, because the original area texts are often short.

This repository contains the human-annotated test set. The LLM-judged training data is published separately as IntelliProcure/TENDRIL-Train.

Dataset structure

The benchmark is split into one retrieval task per tender project. Each project has three configs in MTEB retrieval format, all with the split test:

Config	Columns	Content
`<project>-corpus`	`id`, `text`	Passages from the project's tender documents
`<project>-queries`	`id`, `text`	Areas of action annotated in at least one of the project's documents
`<project>-qrels`	`query-id`, `corpus-id`, `score`	Binary relevance (1 = relevant, 0 = not relevant); unjudged pairs count as not relevant

Project names are anonymized aliases of the form <sector>-<hash>, for example Food-60cd3a. The hash is a keyed hash of the project and stays the same across releases. Query IDs have the form <catalog>::HF-<nnn>. mteb_dataset.json lists all projects.

	Count
Projects (tasks)	33
Sectors	14
Documents	139
Project-query pairs	469
Distinct queries	208
Passages	4,845 (1,557 relevant spans, 3,288 other passages)
Qrels	66,771 (1,870 relevant)

Projects per sector: Construction_Area 4, Construction_Buildings 3, Construction_Civil_Engineering_Infrastructure 2, Electricity 4, Food 4, Furniture 2, Green_Spaces 2, Info_Com_Tech 2, Light_Elect_Appl 2, Paper_and_Stationery 3, Print_Services 1, Road_Transport 2, Textiles 1, Wastepaper_Cleaning_Refuse 1.

Usage

Each project is an MTEB retrieval task. Results are reported as the mean over all projects, with each project counting once.

from mteb.abstasks.retrieval_dataset_loaders import RetrievalDatasetLoader

split = RetrievalDatasetLoader(
    hf_repo="IntelliProcure/TENDRIL", revision="main", split="test", config="Food-60cd3a"
).load()

To load a single component with datasets:

from datasets import load_dataset

corpus = load_dataset("IntelliProcure/TENDRIL", "Food-60cd3a-corpus", split="test")

Dataset creation

Source documents. Tender documents of Swiss public procurement procedures published on simap.ch. Passages were extracted from the PDF and Office attachments with the IntelliProcure document pipeline. Most documents are German; some are French or English.

Annotation. Seven annotators with expertise in sustainability or public procurement marked text spans in the tender documents that address an area of action or one of its criteria, using INCEpTION and written guidelines. Each catalog was annotated by at least two annotators. Two annotators then discussed their disagreements and agreed on a curated version, which is used as the gold standard. Only documents whose curation is marked as finished are included. Criterion annotations are mapped to their parent area of action. Spans that overlap or lie within five characters of each other are merged.

Corpus construction. Every curated span becomes a relevant passage for its areas of action. All other pipeline passages of the same documents that do not overlap a relevant span are added as non-relevant passages. Documents without any curated span are left out, because an empty document does not necessarily indicate the absence of relevant content, for example when it was out of scope or a duplicate of a document annotated in another project. Every query of a project is judged against every non-relevant passage of that project, while relevant spans are only judged for the areas of action they were annotated with. Pairs without a qrel count as not relevant, as in standard MTEB evaluation.

Inter-annotator agreement. Agreement was computed per annotator pair on the documents both annotated, before curation. Passages were matched one-to-one at an intersection over union of at least 0.1, and the sets of areas of action of matched passages were compared. Krippendorff's α with the MASI distance has a median of about 0.48 over 11 pairs, and 0.44–0.56 for the five pairs with the most matched passages. Agreement is moderate, as expected for long documents with many partly overlapping areas of action.

Restricted documents. Projects whose tender documents limit their use to the preparation of an offer, or prohibit publication or disclosure to third parties, were identified with keyword rules plus an LLM check, and excluded. Of the 46 annotated projects, 13 were removed for this reason.

Personal data. Passages were anonymized with gheim: person names, e-mail addresses, phone numbers, URLs, account numbers and credentials are replaced by placeholders such as <PERSON_1>. Placeholders are consistent within a document. Dates and postal addresses are kept. Automatic anonymization can miss entities; please report any personal data you find.

Project identity. Project names are aliases. The mapping to SIMAP project numbers is not published.

Access and versions. The dataset is gated: access is granted automatically after accepting the license terms. Passages are removed on request, for example when a person reports remaining personal data. Each release is marked with a git tag, so that earlier versions of the benchmark stay available; pin a tag or commit with revision=.

Limitations

Small sectors. Several sectors have only one or two projects, so per-sector scores have wide confidence intervals.
Binary labels. Relevance is binary, and passages are spans of variable length, so relevant spans can be much shorter than pipeline passages.
Swiss context. Queries use Swiss German spelling conventions and Swiss procurement terminology.

License

The annotations, queries, selection and anonymization (the Licensed Contributions) are released under the TENDRIL Open RAIL-D License, which permits commercial use but adds use-based restrictions, such as a ban on re-identifying persons and a duty to disclose training on the test data. The tender documents themselves remain the property of their authors; see the license for details.

Citation

@misc{rolshoven2026tendril,
  title  = {{TENDRIL}: Benchmarking Sustainability-Focused Retrieval in Swiss Public Procurement},
  author = {Rolshoven, Luca and Probst, Sophie and Matoshi, Veton and Ellendorff, Tilia and Meili, Rahel and
            Hostettler, Sarah and Jann, Jeremy Austin and Jans, Meret and Tschumi, Pascal and St{\"u}rmer, Matthias},
  year   = {2026}
}

Downloads last month: 118

Total file size:

2.25 MB

Collection including IntelliProcure/TENDRIL

TENDRIL Datasets

Collection

Datasets for training and testing Information Retrieval systems on finding sustainability-related passages in Swiss public call for tender docs. • 2 items • Updated 2 days ago