S

Splade Disco Human Mistral

Developed by slupart
A conversational search model improved based on SPLADE++, optimized for multi-turn dialogue query semantic understanding through multi-teacher distillation strategy
Downloads 27
Release Time : 4/17/2025

Model Overview

This model is a sparse retrieval model optimized for conversational search, retaining the original SPLADE++ document encoder and fine-tuning the query encoder on the QReCC dataset, effectively handling multi-turn conversational search scenarios.

Model Features

Multi-teacher Knowledge Distillation
Combines human annotations and Mistral large model rewritten versions for distillation training, enhancing conversational query understanding capabilities
Dialogue History Processing
Supports flattened dialogue history sequence input, integrating multi-turn dialogue context through [SEP] separators
Asymmetric Architecture
Query encoder and document encoder can be used independently, supporting different representation model combinations

Model Capabilities

Conversational Query Understanding
Multi-turn Context Retrieval
Sparse Vector Generation
Semantic Expansion Retrieval

Use Cases

Conversational Search Systems
Multi-turn Q&A Systems
Handles continuous Q&A scenarios with contextual dependencies
Better understands dialogue context compared to traditional retrieval models
Customer Service Bots
Provides accurate knowledge base retrieval based on dialogue history
Reduces the need for users to repeatedly explain their needs
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase