🤖 AI Summary
Papr has unveiled the papr-embed-code-1, a code-tuned embedding model that refines the existing Qwen3-Embedding for enhanced code retrieval tasks. Available in two sizes (0.6B and 4B), this model has achieved top scores on the CoIR benchmark and ranked first on MTEB leaderboards, showcasing its efficacy. Notably, the 4B variant demonstrated significant improvements in retrieving code it was not specifically trained on, surpassing its base model by an average of 0.03 nDCG@10 across various testing scenarios. However, performance varied, particularly in prose searches, where it fell short of the base model.
This development is significant for the AI/ML community as it underscores the potential benefits of specialized models for niche tasks like code retrieval. The 4B model, when compressed using 1-bit TurboQuant, achieved a score that exceeded the 0.6B model while using only a fraction of the RAM, enhancing efficiency without sacrificing performance. Moreover, the introduction of a hybrid BM25 tokenizer allows for the integration of keyword searches, although dense search remains superior in most code-related tasks. Papr has released these models with open weights, encouraging further exploration and innovation in code retrieval strategies.
Loading comments...
login to comment
loading comments...
no comments yet