comp4cls/comp4cls-4B

Comp4Cls: Semantic Compression for Enhanced Classification

Comp4Cls is a 4 billion parameter model developed by Chanuk Lim, focusing on retrieval-augmented classification (RAG) for scientific and technical documents. Its core innovation lies in entity-centric semantic compression, which converts lengthy texts into short, structured summaries while preserving critical discriminative signals.

Key Capabilities & Features

• Efficient Classification: A 4B-scale model that achieves or surpasses the performance of 8B–14B models, especially in fine-grained categories, by operating on compressed texts.
• Semantic Compression: Uses a two-stage prompting process (entity extraction → selective rewriting) to create concise summaries with an explicit compression ratio, reducing input tokens by ~50% on average.
• RAG with Short Contexts: Operates on compressed texts for both queries and retrieved neighbors, mitigating "lost-in-the-middle" issues and allowing for broader top-k retrieval.
• Scalability & Robustness: Evaluated on large, bilingual datasets including papers, patents, and R&D reports, demonstrating robustness across domains and hierarchical, multi-label taxonomies.
• Production-Ready: Designed for low-latency and high-throughput deployment, with compressed outputs ready for standard vector databases and supporting downstream tasks like semantic search and TL;DR summarization.

Good For

• Classifying scientific and technical documents (e.g., patents, research papers, R&D reports).
• Reducing computational costs and latency in document classification pipelines.
• Implementing efficient retrieval-augmented generation (RAG) systems where context length is a concern.
• Building knowledge organization systems that require structured, compressed representations of documents.