cometadata/ms-marco-ror-reranker

ms-marco-ror-reranker

A cross-encoder reranker fine-tuned for Research Organization Registry (ROR) affiliation matching.

Model Description

This model is fine-tuned from cross-encoder/ms-marco-MiniLM-L-12-v2 on ROR affiliation matching data. It reranks candidate ROR organizations given an affiliation string query.

Training

• —Base model: cross-encoder/ms-marco-MiniLM-L-12-v2
• —Training examples: 45,061
• —Training traces: 2,004
• —Negative sampling: Hard negatives from retrieval candidates
• —Epochs: 3
• —Batch size: 16
• —Learning rate: 2e-05
• —Max sequence length: 256

Usage

from sentence_transformers import CrossEncoder

model = CrossEncoder("cometadata/ms-marco-ror-reranker")

# Score affiliation-candidate pairs
pairs = [
    ["University of California, Berkeley", "University of California, Berkeley"],
    ["University of California, Berkeley", "University of California, Los Angeles"],
]
scores = model.predict(pairs)
print(scores)  # Higher score = better match

Intended Use

This model is designed for reranking ROR organization candidates in affiliation matching pipelines. It should be used after an initial retrieval step (e.g., dense retrieval with Snowflake Arctic).

Training Data

Trained on traces from cometadata/ror-pipeline-traces (affrodbs2afftraces config).

Timestamp

2026-01-07T21:35:26.376404+00:00