What is Qwen3 Reranker
Qwen3 Reranker is a text reranking model released by Alibaba's Tongyi Qianwen team, belonging to theQwen3model family. It adopts a single-tower cross-encoder architecture, outputting a relevance score after inputting text pairs. The model is trained through a multi-stage training paradigm based on high-quality annotated data and a large number of synthetic training pairs, supporting over 100 languages, covering mainstream natural languages and multiple programming languages. In terms of performance, Qwen3 Reranker-8B achieved a high score of 72.94 on the MTEB leaderboard, and Qwen3 Reranker-0.6B has also surpassed Jina and BGE Rerankers.
Main Features of Qwen3 Reranker
- Text relevance assessment: Qwen3 Reranker can accept text pairs such as user queries and candidate documents as input. The model calculates and outputs the relevance score between two texts; the higher the score, the stronger the relevance between the text pair.
- Text reranking: Based on the relevance score, Qwen3 Reranker can rank candidate texts, placing the texts most relevant to the user query first.
-
Optimize retrieval results: In semantic retrieval scenarios, this feature can help users find the most relevant information faster, improving retrieval efficiency and accuracy.
-
Support multiple languages: Qwen3 Reranker supports over 100 languages, including mainstream natural languages and multiple programming languages.
Technical Principles of Qwen3 Reranker
-
Single-tower cross-encoder architecture: Qwen3 Reranker adopts a single-tower cross-encoder architecture, concatenating the query and candidate document and inputting them together into the model. The architecture can process text pairs simultaneously, perform deep interaction analysis, and more accurately assess the relevance between text pairs.
-
Instruction-aware capability: The model supports dynamically defining task objectives through instructions, such as "determine the relevance of legal documents," enabling the model to flexibly adjust relevance assessment criteria according to different tasks and scenarios.
-
Input and output format: The input format is a specific chat template, containing instructions, queries, and documents. The model outputs a relevance score, determined by calculating the probability of "yes" or "no."
-
Multi-stage training paradigm: The training process of Qwen3 Reranker includes multiple stages. In the supervised fine-tuning stage, high-quality annotated data is used directly for training, improving training efficiency. Synthetic data is used to further enhance model performance.
-
High-quality data filtering: In the selection of training data, Qwen3 Reranker uses a variety of high-quality annotated datasets, such as MS MARCO, NQ, HotpotQA, etc. High-quality synthetic data pairs are filtered out through cosine similarity.
-
Model fusion technology: Spherical linear interpolation (SLERP) technology is used to fuse parameters of multiple model checkpoints saved during the fine-tuning stage. It can better preserve the geometric characteristics of the model parameters and improve the model's robustness and generalization performance across different data distributions.
-
Supervised fine-tuning loss (SFT Loss): Qwen3 Reranker optimizes the supervised fine-tuning loss function. By maximizing the probability of the correct label ("yes" or "no"), the model learns to distinguish relevant from irrelevant documents. This binary classification approach simplifies the reranking task, based on the instruction-following capability of the Qwen3 model.
Project Address of Qwen3 Reranker
- HuggingFace Model Library：https://huggingface.co/collections/Qwen/qwen3-reranker
Application Scenarios of Qwen3 Reranker
-
Semantic search: In scenarios such as search engines and question-answering systems, rerank search results so that the most relevant content is displayed first.
-
Text classification: By evaluating the relevance between text and category labels, assist in text classification tasks and improve classification accuracy.
-
Sentiment analysis: In sentiment analysis, rank the relevance between reviews or text and sentiment labels to assist in determining sentiment orientation.
-
Code search: In code repositories, rank the relevance between code snippets and user queries to help developers quickly find relevant code.