cometadata/jina-reranker-v2-multilingual-affiliations-comet-training-only

Jina Affiliation Reranker

Cross Encoder reranker model fine-tuned for affiliation string matching. Given a pair of affiliation strings, it predicts how likely they refer to the same institution.

Use Case

This model is designed for matching and disambiguating messy real-world affiliation strings against canonical institution records (ROR).

Examples of what it handles:

• Abbreviations: "MIT" ↔ "Massachusetts Institute of Technology"
• Word reordering: "University of Oxford" ↔ "Oxford University"
• Partial matches: "Dept. of Physics, Stanford" ↔ "Stanford University"
• International variants: "東京大学" ↔ "University of Tokyo"
• OCR noise: "Univ ersity of Cal ifornia" ↔ "University of California"

Usage

from sentence_transformers import CrossEncoder

model = CrossEncoder(
    "cometadata/jina-reranker-v2-multilingual-affiliations",
    trust_remote_code=True,
)

# Score affiliation pairs (higher = more likely same institution)
pairs = [
    ["University of California, Berkeley", "UC Berkeley"],
    ["University of California, Berkeley", "Berkeley College"],
]
scores = model.predict(pairs)
# [0.82, 0.15] - first pair matches, second doesn't

# Rank candidates for an affiliation string
results = model.rank(
    "MIT, Cambridge, MA",
    [
        "Massachusetts Institute of Technology",
        "MIT University (India)",
        "University of Cambridge",
    ]
)
# Returns candidates ranked by relevance

Training

Base Model: jinaai/jina-reranker-v2-base-multilingual

Dataset: cometadata/triplet-loss-for-embedding-affiliations-sample-1

• ~8K triplets (anchor, positive, negative)
• 80% hard negatives (similar but different institutions)
• 20% easy negatives (clearly different institutions)

Configuration:

Parameter	Value
Epochs	3
Batch size	16
Learning rate	2e-5
Loss	BinaryCrossEntropyLoss
Validation split	15%

Evaluation

Evaluated on 300 test cases across 10 difficulty tiers:

Tier	Cases	Base Model	Fine-tuned	Δ
Baseline	30	100.0%	100.0%	—
OCR/Noise	30	100.0%	100.0%	—
Abbreviations	40	60.0%	80.0%	+20.0%
Hierarchical	35	71.4%	77.1%	+5.7%
Medical/Hospital	25	64.0%	68.0%	+4.0%
Research Labs	25	80.0%	84.0%	+4.0%
International	35	82.9%	91.4%	+8.6%
Disambiguation	31	45.2%	51.6%	+6.5%
Negative Controls	19	100.0%	100.0%	—
Ultra-Hard	30	93.3%	96.7%	+3.3%

Overall: 78.3% → 84.3% accuracy (+6.0%), MRR 0.873 → 0.913

Model Details

• Parameters: 278M
• Max sequence length: 1024 tokens
• Output: Single relevance score (0-1)
• Languages: Multilingual (inherits from base model)

License

CC-BY-NC-4.0 (inherited from base model - non-commercial use only)

Citation

@misc{jina-affiliation-reranker,
  title={Jina Affiliation Reranker},
  author={cometadata},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/cometadata/jina-reranker-v2-multilingual-affiliations}
}