IntelliProcure/TENDRIL-most-embed-de
Sentence Similarity • 1B • Updated • 7
Datasets:
Tasks:
Text Retrieval
Modalities:
Text
Formats:
parquet
Sub-tasks:
document-retrieval
Size:
100K - 1M
License:
You need to agree to the conditions to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
TENDRIL-Train 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.
Log in or Sign Up to review the conditions and access this dataset content.
TENDRIL-Train: LLM-Judged Tender Retrieval Data
Training data for retrieving sustainability-relevant passages from Swiss public procurement tenders. It accompanies the human-annotated benchmark IntelliProcure/TENDRIL. Queries are the areas of action (Handlungsfelder) and criteria of our sustainability catalogs, which are based on Swiss and EU recommendations for sustainable public procurement. Each query-passage pair was graded by a large language model.
Dataset structure
The configs are in MTEB retrieval format, all with the split train:
| Config | Columns | Rows |
|---|---|---|
corpus |
id, text, project, document, page |
85,772 passages |
queries |
id, text, level, catalog, area |
1,041 queries |
qrels (default) |
query-id, corpus-id, score |
911,540 judgments |
score: 0 = not relevant, 1 = partially relevant, 2 = relevant.level:area_of_action(413 queries) orcriterion(628 queries). For criteria,areagives the parent area of action.projectanddocument: keyed hashes of the SIMAP project number and of the document file name. They let you group passages by project or document (631 projects, 2,892 documents) but cannot be reversed. A passage was only judged against queries selected for its own project.page: the page of the source document the passage comes from.
| Level | Score 0 | Score 1 | Score 2 |
|---|---|---|---|
| Area of action | 413,877 | 30,631 | 12,141 |
| Criterion | 425,928 | 22,570 | 6,393 |
| Total | 839,805 | 53,201 | 18,534 |
Queries per catalog: Construction_Buildings 152, Construction_Area 127, Light_Elect_Appl 126, Construction_Civil_Engineering_Infrastructure 124, Info_Com_Tech 123, Road_Transport 114, Furniture 63, Product_Independent_Criteria 61, Print_Services 48, Green_Spaces 41, Food 26, Paper_and_Stationery 16, Electricity 12, IT_Software 8.
Each row covers the judgments against one catalog's queries, with the number of projects in parentheses. A project can be judged against several catalogs, so projects, documents, and passages repeat across rows, and the column headers give the unique totals. A positive is a judgment with score 1 or 2.
from datasets import load_dataset
corpus = load_dataset("IntelliProcure/TENDRIL-Train", "corpus", split="train")
queries = load_dataset("IntelliProcure/TENDRIL-Train", "queries", split="train")
qrels = load_dataset("IntelliProcure/TENDRIL-Train", "qrels", split="train")
Dataset creation
- Project and document selection. 800 tender projects published on simap.ch were sampled, with up to five documents each. Every project gets one criterion per selected area of action. Areas are drawn by quota from tiers of prior relevance evidence (validated relevant, unvalidated relevant, unsure, irrelevant), plus a coverage top-up so that every area appears in at least three projects. Projects of the TENDRIL test set are excluded. The random seed is 42.
- Passage filtering. Passages from the IntelliProcure document pipeline are kept if they have 20–5,000 characters, are not dominated by digits or symbols, are not degenerate tables, and are not near-duplicates (5-gram Jaccard ≥ 0.7).
- Anonymization. Person names, e-mail addresses, phone numbers, URLs, account numbers and credentials are
replaced by placeholders such as
<PERSON_1>, using gheim. - Query wording. Areas of action use a German summary of at most 100 words, generated from the catalog by
deepseek/deepseek-v4-flash(temperature 0). Criteria use the criterion text followed by its ambition levels (Basis, Gute Praxis, Vorbild). The same wording is used in the test set. - Relevance judgments. Each query-passage pair of a project was graded by
z-ai/glm-5.3-flash(low reasoning effort, temperature 0) via OpenRouter, using the shortest of three versions of the German annotation guidelines, in October 2026. Of 906,984 requests, 7 could not be parsed and 47 were truncated. - Restricted documents. Projects whose tender documents limit their use to the preparation of an offer, or prohibit publication or disclosure to third parties, were excluded (169 of the 800 projects).
Label quality
The judge was selected on a human-annotated calibration set of 1,425 query-passage pairs (three annotators, ten catalogs), out of eight LLMs at two reasoning efforts and three prompt versions. Agreement between the selected judge and the curated human labels:
| Level | Cohen's κ (0/1/2) | Cohen's κ (relevant vs. not) |
|---|---|---|
| Area of action | 0.48 | 0.62 |
| Criterion | 0.53 | 0.65 |
For comparison, Krippendorff's α between the three human annotators is 0.57 (area) and 0.59 (criterion). The labels are noisy and should be treated as weak supervision.
Limitations
- Missing test sectors. No training queries come from the Textiles or Wastepaper_Cleaning_Refuse catalogs, which appear in the test set.
- Skewed labels. Queries have very different numbers of judgments (median 349, maximum 22,170), and 92% of judgments are 0.
- Imperfect anonymization. Automatic anonymization can miss entities; please report any personal data you find.
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; pin a tag
or commit with revision=.
License
The selection, queries, relevance judgments, anonymization and identifiers (the Licensed Contributions) are released under the TENDRIL Open RAIL-D License, which permits commercial use but adds use-based restrictions specific to procurement. Models trained on this data are not Derivatives of the Data under that license. 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
- 33