maayans/modernbert-embed-large__ratio-address

modernbert-embed-large__ratio-address

lightonai/modernbert-embed-large fine-tuned for the ADDRESS ideation operation of RATIO (Retrieval Across Typed Ideation Operations): given a scientific query sentence, retrieve an approach or insight that responds to a problem stated in the query.

One of 9 checkpoints in the RATIO release (3 encoders x ADDRESS / BROADEN / SPECIFY); see the dataset card for the benchmark.

Input format (important)

This checkpoint was trained with literal text prefixes: search_query: for queries, search_document: for candidate sentences. Use the same format at inference:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("maayans/modernbert-embed-large__ratio-address")
q_emb = model.encode(["search_query: " + q for q in queries])
d_emb = model.encode(["search_document: " + d for d in candidates])

Training

Relation-specific contrastive fine-tuning (MultipleNegativesRankingLoss) on the RATIO address train split. The benchmark uses a temporal split: all test papers postdate the training data of every evaluated encoder. See the paper for details.

Citation

@misc{sharon2026ratiobenchmarkretrievaltyped,
      title={RATIO: A Benchmark for Retrieval Across Typed Ideation Operations in Scientific Literature}, 
      author={Maayan Sharon and Tom Hope},
      year={2026},
      eprint={2608.27394},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.27394}, 
}