ai4data/devdata-search-multilingual-e5-small-cmnrl

1---2license: apache-2.03library_name: sentence-transformers4pipeline_tag: feature-extraction5tags:6  - sentence-transformers7  - feature-extraction8  - retrieval9  - devdata-search10datasets:11  - ai4data/devdatabench12base_model: intfloat/multilingual-e5-small13---14 15# devdata-search-multilingual-e5-small-cmnrl16 17A bi-encoder embedding model for **search over structured statistical18metadata**, part of the **DevData Search** family. It is a fine-tune of19`intfloat/multilingual-e5-small` produced with schema-invariant fine-tuning on20[DevDataBench](https://huggingface.co/datasets/ai4data/devdatabench): full-schema21serialization with per-example field-order permutation and field dropout, so the22encoder binds meaning to field labels rather than to serialization order. This is23an embedding model that powers retrieval; it is not a hosted search service.24 25See the paper *Field Order Should Not Matter: Permutation-Invariant Fine-Tuning26for Structured Metadata Retrieval*.27 28## Training29 30- Base model: `intfloat/multilingual-e5-small`31- Loss: `cmnrl`32- Field permutation: `True`; field dropout: `0.15`33- Max sequence length: `512`34- Query prefix: `query: ` ; document prefix: `passage: ` (prepend these when encoding)35 36## Usage37 38```python39from sentence_transformers import SentenceTransformer40 41model = SentenceTransformer("ai4data/devdata-search-multilingual-e5-small-cmnrl")42queries = ["query: " + "mobile-broadband subscriptions per 100 people"]43docs = ["passage: " + "name: Active mobile-broadband subscriptions | ..."]44q = model.encode(queries)45d = model.encode(docs)46```47 48Cosine similarity of `q` and `d` ranks documents for each query.49 50## License51 52Apache-2.0. Derived from `intfloat/multilingual-e5-small`; trained on public World Bank Data360 metadata.53