castorini/first_qwen3_14b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:14BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 25, 2025Architecture:Transformer Featherless Exclusive Cold

The castorini/first_qwen3_14b is a 14 billion parameter language model developed by Castorini, specifically fine-tuned for reranking retrieval results. It excels in information retrieval tasks, particularly for improving the nDCG@10 scores on DL tracks. This model is designed to enhance the relevance of search results by processing the top 100 outputs from a first-stage retriever like SPLADE++ EnsemBleDistil.

Loading preview...

Model Overview

The castorini/first_qwen3_14b is a 14 billion parameter model developed by Castorini, specifically fine-tuned for reranking in information retrieval systems. It is designed to improve the quality of search results by re-evaluating the top 100 documents initially retrieved by a first-stage retriever, such as SPLADE++ EnsemBleDistil.

Key Capabilities

  • Enhanced Reranking Performance: Demonstrates strong performance in reranking tasks, significantly improving nDCG@10 scores across various DL tracks (DL19, DL20, DL21, DL22, DL23).
  • Information Retrieval Optimization: Optimized for refining retrieval results, making it suitable for applications requiring high precision in document ranking.
  • Comparative Performance: Outperforms several zero-shot Qwen3 models of varying sizes (0.6B, 1.7B, 4B, 8B, 14B, 32B) in reranking benchmarks, showcasing the effectiveness of its fine-tuning.

Use Cases

  • Search Engines: Ideal for integrating into search pipelines to boost the relevance of search results.
  • Question Answering Systems: Can be used to rerank retrieved passages to provide more accurate answers.
  • Document Retrieval: Applicable in scenarios where precise ordering of documents based on relevance is critical.