Embeddings for Spanish public procurement
A semantic encoder fine-tuned on historical versions of Spanish procurement notices and text-to-CPV-description pairs. It is intended for search, related-notice retrieval, version matching, and CPV division retrieval.
It was trained from the hsilvosa/openplacsp.
Temporal evaluation on 2024
| Task | Metric | Base | Fine-tuned |
|---|---|---|---|
| Retrieve another version | Recall@1 | 0.9984 | 0.9982 |
| Retrieve another version | Recall@10 | 1.0000 | 0.9998 |
| Retrieve a CPV division | Recall@1 | 0.1920 | 0.6846 |
| Retrieve a CPV division | Recall@3 | 0.3672 | 0.8550 |
Training only uses notices whose first publication date is no later than 2022. The years
2023 and 2024 are reserved for validation and testing. Training uses
150,000 version pairs and
50,000 text-to-CPV pairs with seed
20260817. The source snapshot fingerprint is
fad46713c99abcaa500c7cef9323ae173f8f75c5aca02b299db2d4e96c3ca934 and the model checksum is e89f3596034c54b46d3959d2c7a33e378c9d67019405c9b6d26e56f207ec01ed.
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer(".")
queries = model.encode(["query: mantenimiento de aplicaciones"], normalize_embeddings=True)
documents = model.encode(["passage: servicios de desarrollo de software"], normalize_embeddings=True)
similarity = model.similarity(queries, documents)
Limitations
The model reflects Spanish administrative language and data published through December 2024. Similarity does not imply legal identity, irregularity, or contractual equivalence. CPV descriptions are short and some divisions have relatively few examples.