Salesforce/SweRankLLM-Large
Salesforce/SweRankLLM-Large is a 32.8 billion parameter language model based on Qwen2.5-32B-Instruct, fine-tuned for listwise code-reranking. With a 32768 token context length, it significantly enhances the quality of results for software issue localization when combined with performant code retrievers. The model was trained on large-scale issue localization data from public Python GitHub repositories, making it specialized for improving code search and relevance.
Loading preview...
SweRankLLM-Large: Specialized for Code Reranking
SweRankLLM-Large is a 32.8 billion parameter language model developed by Salesforce, built upon the Qwen2.5-32B-Instruct architecture. Its primary distinction lies in its fine-tuning for listwise code-reranking, a process designed to improve the relevance and quality of results in software issue localization.
Key Capabilities & Features
- Code Reranking: Specifically optimized to re-order lists of code snippets, enhancing the precision of search results.
- Software Issue Localization: When paired with effective code retrievers like SweRankEmbed, it significantly boosts the accuracy of identifying relevant code for software issues.
- Large Context Window: Features a substantial 32768 token context length, allowing it to process and understand longer code sequences and contextual information.
- Specialized Training Data: Trained on extensive issue localization data sourced from public Python GitHub repositories, ensuring its proficiency in real-world software development contexts.
Use Cases & Benefits
- Improved Code Search: Ideal for applications requiring highly relevant code search results, particularly in debugging and issue resolution.
- Enhanced Developer Productivity: By providing more accurate code suggestions and localizations, it can streamline the process of identifying and fixing software bugs.
- Research in Software Engineering: Useful for researchers exploring advanced techniques in code understanding, retrieval, and ranking. More details are available in the associated blog post and paper.