TENDRIL Arctic-Embed-L v2.0

A fine-tuned version of Snowflake/snowflake-arctic-embed-l-v2.0 for retrieving sustainability-relevant passages from Swiss public procurement tenders. Queries describe areas of action or criteria from our sustainability catalogs, which are based on Swiss and EU recommendations for sustainable public procurement.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("IntelliProcure/TENDRIL-arctic-embed-l-v2.0")
queries = model.encode_query(["Regionale und saisonale Lebensmittel bevorzugen ..."])
passages = model.encode_document(["Die Anbieterin liefert vorwiegend Produkte aus der Region ..."])
scores = model.similarity(queries, passages)

encode_query adds the prompt query: ; documents get no prompt. Embeddings use CLS pooling, are L2-normalized and have 1,024 dimensions, and inputs are truncated at 8,192 tokens. Truncating to 256 dimensions (Matryoshka) was supported by the base model but was not evaluated after fine-tuning.

Results

Test set: IntelliProcure/TENDRIL (revision 15eef0bc), 33 projects. Scores are macro-averaged over projects, and 95% intervals come from 10,000 project-level bootstrap resamples.

Model	nDCG@10	Recall@10	Recall@100	MRR@10
BM25	0.421 [0.381, 0.462]	0.511	0.929	0.463
snowflake-arctic-embed-l-v2.0 (base)	0.547 [0.509, 0.585]	0.676	0.966	0.573
TENDRIL Arctic-Embed-L v2.0	0.642 [0.596, 0.690]	0.778	0.990	0.652

Fine-tuning improves nDCG@10 by +0.095 [+0.063, +0.130] in a two-sided paired bootstrap over projects; the gain is significant for all reported metrics after a Holm correction over the four fine-tuned models in the paper (p < 0.01).

On the validation projects, balanced nDCG@10 (the mean over areas of action and criteria) rose from 0.620 to 0.823.

Training

Data. Training examples were built from the LLM-judged TENDRIL-Train data: a query, one relevant passage and three hard negatives from the same project. Area-of-action and criterion examples were balanced, giving 9,000 of 13,390 available examples. Validation used held-out projects (63 for areas of action, 62 for criteria).
Loss. InfoNCE over the positive and the three negatives, with the cosine similarity of L2-normalized embeddings and a similarity scale of 20. Positives are passages the judge rated relevant (2); negatives are passages rated irrelevant (0) for the same query, two from the positive's document and one from another document of the same project (the other pool fills in when one has too few).
Hyperparameters. Learning rate 1e-5 with linear decay and 10% warmup, AdamW without weight decay, effective batch size 32 (1 × 32 gradient accumulation), at most 2 epochs, bf16, maximum sequence length 8,192, seed 42. The model was evaluated every 56 steps, and the checkpoint with the best balanced validation nDCG@10 was kept.
Software. sentence-transformers 6.0.1, transformers 5.17.0, PyTorch 2.6.0.

Limitations

German-centred. The model was trained on German queries and mostly German passages; behaviour on French or Italian tenders was not evaluated.
Noisy labels. Training labels come from an LLM judge whose agreement with human annotators is moderate (Cohen's κ 0.48–0.53 on the 0/1/2 scale).
Not a decision tool. Rankings indicate where a tender may address a sustainability topic. They are not an assessment of a tender or a bidder and must not be used for automated procurement decisions.

License

Apache License 2.0, the same as the base model. See LICENSE and NOTICE.

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}
}