You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

RATIO: Retrieval Across Typed Ideation Operations

A benchmark for retrieving scientific ideation moves: given a research statement (query sentence), retrieve sentences from other papers that Address, Broaden, or Specify it. Built from full-text computer-science papers.

Structure

data/
  address/     train.parquet  validation.parquet  test.parquet  silver.parquet   # query-gold pairs + judge-validated test subset
  broaden/     train.parquet  validation.parquet  test.parquet  silver.parquet   # query-gold pairs + judge-validated test subset
  specify/     train.parquet  validation.parquet  test.parquet  silver.parquet   # query-gold pairs + judge-validated test subset
  candidates/  train.parquet  validation.parquet  test.parquet                                                     # shared retrieval corpus
  • Each relation config contains query-gold sentence pairs; every query has exactly one gold.
  • candidates is the retrieval corpus per split (golds + distractors), shared across all three relations.
  • silver is the subset of the test set whose (query, gold) pair was validated by an LLM judge under two independent prompts (both must accept). Silver retrieval uses the test candidates corpus.
  • Splits are temporal: train 2015-Sep 2025, validation Q4 2025, test 2026 onward.

Usage

from datasets import load_dataset

pairs  = load_dataset("maayans/RATIO", "address")     # train / validation / test / silver
corpus = load_dataset("maayans/RATIO", "candidates")  # retrieval corpus (train / validation / test)

Retrieval for a given split is performed against the candidates corpus of that same split; the silver subset shares the test candidates.

Models

Nine fine-tuned bi-encoders accompany the benchmark - three encoders x three relations, named maayans/<base-model>__ratio-<relation> with relation in address / broaden / specify (e.g. maayans/modernbert-embed-large__ratio-address), all grouped in the RATIO collection.

Base model When to use it Query prefix Candidate prefix
modernbert-embed-large Top choice - best overall retrieval quality search_query: search_document:
stella_en_1.5B_v5 Very close performance; preferred for longer texts built-in s2p_query prompt none
all-mpnet-base-v2 When a light and fast model is needed query: document:

Prefixes must be applied at inference exactly as in training:

from sentence_transformers import SentenceTransformer

m = SentenceTransformer("maayans/modernbert-embed-large__ratio-address")
q = m.encode(["search_query: " + s for s in queries])
c = m.encode(["search_document: " + s for s in candidates])

# stella: built-in prompt on queries only, candidates unprefixed
m = SentenceTransformer("maayans/stella_en_1.5B_v5__ratio-address", trust_remote_code=True)
q = m.encode(queries, prompt_name="s2p_query")
c = m.encode(candidates)

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}, 
}
Downloads last month
15

Models trained or fine-tuned on maayans/RATIO

Collection including maayans/RATIO

Paper for maayans/RATIO