Salesforce/SweRankEmbed-Small
Salesforce/SweRankEmbed-Small is a 0.1 billion parameter bi-encoder model developed by Salesforce, specifically designed for code retrieval. It supports an 8192 token context length and is trained on large-scale issue localization data from public Python GitHub repositories. This model excels at issue localization tasks, significantly outperforming other embedding models and agent-based approaches.
Loading preview...
SweRankEmbed-Small: Code Retrieval Bi-Encoder
SweRankEmbed-Small is a compact yet powerful 137 million parameter bi-encoder model from Salesforce, engineered for efficient code retrieval. It boasts an 8192 token context length, making it suitable for processing substantial code snippets.
Key Capabilities
- Code Retrieval: Optimized for finding relevant code based on natural language queries.
- Issue Localization: Demonstrates superior performance in identifying code locations related to specific issues, outperforming larger models and agent-based systems like Claude-3.5 on benchmarks such as SWE-Bench-Lite and LocBench.
- Scalable Training: Trained on extensive issue localization data sourced from public Python GitHub repositories.
- Integration: Can be combined with
SweRankLLM-SmallorSweRankLLM-Largererankers for enhanced ranking quality.
Performance Highlights
SweRankEmbed-Small achieves a 74.45% Func@10 on SWE-Bench-Lite and 63.39% Func@15 on LocBench, surpassing agent-based methods and other embedding models in its class. For instance, it significantly outperforms OpenHands (Claude 3.5) and LocAgent (Claude 3.5) in these metrics.
Good for
- Software Development: Assisting developers in quickly locating relevant code sections or identifying problematic areas.
- Code Search Engines: Powering efficient and accurate code search functionalities.
- Automated Debugging: Contributing to systems that automatically pinpoint code issues.
- Research in Code Intelligence: Serving as a strong baseline or component in advanced code understanding systems.