SPLADE_DistilMSE: SPLADEv2 trained with the distillated triplets

Training data from: https://github.com/sebastian-hofstaetter/neural-ranking-kd

From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective (Thibault Formal, Carlos Lassance, Benjamin Piwowarski, Stéphane Clinchant). 2022. https://arxiv.org/abs/2205.04733

Using the model

The model can be loaded with experimaestro IR

from xpmir.models import AutoModel

# Model that can be re-used in experiments
model, init_tasks = AutoModel.load_from_hf_hub("xpmir/SPLADE_DistilMSE")

# Use this if you want to actually use the model
model = AutoModel.load_from_hf_hub("xpmir/SPLADE_DistilMSE", as_instance=True)
model.rsv("walgreens store sales average", "The average Walgreens salary ranges...")

Results

Dataset	AP	P@20	RR	RR@10	nDCG	nDCG@10	nDCG@20
msmarco_dev	0.3642	0.0382	0.3693	0.3582	0.4879	0.4222	0.4458
trec2019	0.4896	0.7209	0.9496	0.9496	0.7253	0.7055	0.6926
trec2020	0.5026	0.6315	0.9483	0.9475	0.7273	0.6868	0.6627