devdata-search-granite-97m-multilingual-cmnrl
A bi-encoder embedding model for search over structured statistical
metadata, part of the DevData Search family. It is a fine-tune of
ibm-granite/granite-embedding-97m-multilingual-r2 produced with schema-invariant fine-tuning on
DevDataBench: full-schema
serialization with per-example field-order permutation and field dropout, so the
encoder binds meaning to field labels rather than to serialization order. This is
an embedding model that powers retrieval; it is not a hosted search service.
See the paper Field Order Should Not Matter: Permutation-Invariant Fine-Tuning for Structured Metadata Retrieval.
Training
- Base model:
ibm-granite/granite-embedding-97m-multilingual-r2 - Loss:
cmnrl - Field permutation:
True; field dropout:0.15 - Max sequence length:
512 - No query/document prefixes
Usage
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("ai4data/devdata-search-granite-97m-multilingual-cmnrl")
queries = ["mobile-broadband subscriptions per 100 people, reported annually"]
docs = ["name: Active mobile-broadband subscriptions | ..."]
q = model.encode(queries)
d = model.encode(docs)
Cosine similarity of q and d ranks documents for each query.
License
Apache-2.0. Derived from ibm-granite/granite-embedding-97m-multilingual-r2; trained on public World Bank Data360 metadata.