cross-encoder-MiniLM-L12-Hinge

This model is a cross-encoder based on microsoft/MiniLM-L12-H384-uncased. It was trained on Ms-Marco using loss hingeLoss as part of a reproducibility paper for training cross encoders: "Reproducing and Comparing Distillation Techniques for Cross-Encoders", see the paper for more details.

Contents

• Model Description
• Usage
• Evals

Model Description

This model is intended for re-ranking the top results returned by a retrieval system (like BM25, Bi-Encoders or SPLADE).

• Training Data: MS MARCO Passage
• Language: English
• Loss hingeLoss

Training can be easily reproduced using the assiciated repository. The exact training configuration used for this model is also detailed in config.yaml.

Usage

Quick Start:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("microsoft/MiniLM-L12-H384-uncased")
model = AutoModelForSequenceClassification.from_pretrained("xpmir/cross-encoder-MiniLM-L12-Hinge")

features = tokenizer("What is experimaestro ?", "Experimaestro is a powerful framework for ML experiments management...", padding=True, truncation=True, return_tensors="pt")

model.eval()
with torch.no_grad():
    scores = model(**features).logits
    print(scores)

Evaluations

We provide evaluations of this cross-encoder re-ranking the top 1000 documents retrieved by naver/splade-v3-distilbert.

dataset	RR@10	nDCG@10
msmarco_dev	38.68	45.16
trec2019	97.67	73.42
trec2020	95.06	73.72
fever	78.87	79.00
arguana	22.46	33.27
climate_fever	26.81	20.05
dbpedia	74.03	43.09
fiqa	44.61	36.41
hotpotqa	85.90	68.09
nfcorpus	56.50	33.72
nq	51.79	56.76
quora	68.98	72.31
scidocs	27.61	15.34
scifact	67.59	70.06
touche	65.41	33.09
trec_covid	89.35	69.05
robust04	71.52	49.26
lotte_writing	66.06	57.90
lotte_recreation	61.23	55.32
lotte_science	45.44	37.67
lotte_technology	52.88	44.83
lotte_lifestyle	71.60	62.20
Mean In Domain	77.14	64.10
BEIR 13	58.45	48.48
LoTTE (OOD)	61.46	51.20