---
title: naver/splade-v3
description: This checkpoint corresponds to a model that starts from SPLADE++SelfDistil (`naver/splade-cocondenser-selfdistil`), and is trained with a mi. BERT, 110M parameters.
canonical_url: https://superlinked.com/models/naver-splade-v3
last_updated: 2026-05-25
---

# naver/splade-v3

This checkpoint corresponds to a model that starts from SPLADE++SelfDistil (`naver/splade-cocondenser-selfdistil`), and is trained with a mix of KL-Div and MarginMSE, with 8 negatives per query sampled from SPLADE++SelfDistil. We used the original MS MARCO collection without the titles.

Source: [naver/splade-v3 on HuggingFace](https://huggingface.co/naver/splade-v3)

## Overview

| Field | Value |
|-------|-------|
| Architecture | BERT |
| Parameters | 110M |
| Tasks | Encode |
| Outputs | Sparse |
| Dimensions | Sparse: 30,522 |
| Max sequence length | 512 tokens |
| License | cc-by-nc-sa-4.0 |
| Inputs | text |
| Languages | en |

## Benchmarks

### CQADupstackPhysicsRetrieval

Domain: scientific · Task: retrieval · Language: en

Duplicate question retrieval from StackExchange Physics

Corpus: 38,314 · Queries: 1,039

**Quality:** ndcg at 10: 0.3639 · map at 10: 0.3082 · mrr at 10: 0.3560

**Performance (L4 b1 c16):** Corpus 20.7K tok/s · Corpus p50 73.7ms · Query 3.0K tok/s · Query p50 52.9ms

[Reference](http://nlp.cis.unimelb.edu.au/resources/cqadupstack/)

### CosQA

Domain: technology · Task: retrieval · Language: en

Code search with natural language queries

Corpus: 6,267 · Queries: 500

**Quality:** ndcg at 10: 0.2045 · map at 10: 0.1567 · mrr at 10: 0.1702

**Performance (L4 b1 c16):** Corpus 9.1K tok/s · Corpus p50 62.2ms · Query 1.6K tok/s · Query p50 51.3ms

[Reference](https://arxiv.org/abs/2105.13239)

### FiQA2018

Domain: finance · Task: retrieval · Language: en

Financial opinion mining and question answering

Corpus: 57,599 · Queries: 648

**Quality:** ndcg at 10: 0.2768 · map at 10: 0.2113 · mrr at 10: 0.3451

**Performance (L4 b1 c16):** Corpus 24.5K tok/s · Corpus p50 67.6ms · Query 3.3K tok/s · Query p50 51.2ms

[Reference](https://sites.google.com/view/fiqa/)

### LegalBenchConsumerContractsQA

Domain: legal · Task: retrieval · Language: en

Question answering on consumer contracts

Corpus: 153 · Queries: 396

**Quality:** ndcg at 10: 0.7393 · map at 10: 0.6784 · mrr at 10: 0.6805

**Performance (L4 b1 c16):** Corpus 56.0K tok/s · Corpus p50 115.2ms · Query 4.5K tok/s · Query p50 52.2ms

[Reference](https://huggingface.co/datasets/nguha/legalbench)

### NFCorpus

Domain: medical · Task: retrieval · Language: en

Biomedical literature search from NutritionFacts.org

Corpus: 3,593 · Queries: 323

**Quality:** ndcg at 10: 0.3404 · map at 10: 0.1300 · mrr at 10: 0.5417

**Performance (L4 b1 c16):** Corpus 37.0K tok/s · Corpus p50 98.4ms · Query 1.4K tok/s · Query p50 52.6ms

**Performance (RTX-4090 b1 c16):** Corpus 108.8K tok/s · Corpus p50 40.3ms · Query 3.5K tok/s · Query p50 19.1ms

[Reference](https://www.cl.uni-heidelberg.de/statnlpgroup/nfcorpus/)

### SCIDOCS

Domain: scientific · Task: retrieval · Language: en

Citation prediction, document classification, and recommendation for scientific papers

Corpus: 25,656 · Queries: 1,000

**Quality:** ndcg at 10: 0.1543 · map at 10: 0.0878 · mrr at 10: 0.2686

**Performance (L4 b1 c16):** Corpus 26.2K tok/s · Corpus p50 83.4ms · Query 3.0K tok/s · Query p50 53.6ms

[Reference](https://allenai.org/data/scidocs)

### SciFact

Domain: scientific · Task: retrieval · Language: en

Scientific claim verification using research literature

Corpus: 5,183 · Queries: 300

**Quality:** ndcg at 10: 0.6846 · map at 10: 0.6371 · mrr at 10: 0.6524

**Performance (L4 b1 c16):** Corpus 35.8K tok/s · Corpus p50 89.7ms · Query 4.5K tok/s · Query p50 52.1ms

[Reference](https://github.com/allenai/scifact)

### StackOverflowQA

Domain: technology · Task: retrieval · Language: en

Programming question answering from Stack Overflow

Corpus: 19,931 · Queries: 1,994

**Quality:** ndcg at 10: 0.7380 · map at 10: 0.7057 · mrr at 10: 0.7057

**Performance (L4 b1 c16):** Corpus 33.0K tok/s · Corpus p50 84.0ms · Query 50.8K tok/s · Query p50 91.0ms

[Reference](https://arxiv.org/abs/2407.02883)