Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

FineScope: Precision Pruning for Domain-Specialized Large Language Models Using SAE-Guided Self-Data Cultivation

Created by
  • Haebom

Author

Chaitali Bhattacharyya, Hyunsei Lee, Junyoung Lee, Shinhyoung Jang, Il Hong Suh, Yeseong Kim

Outline

FineScope is a framework for deriving compact, domain-specific, optimized LLMs from large-scale pre-trained language models (LLMs) that maintain efficiency and robust performance. FineScope leverages the Sparse Autoencoder (SAE) framework, which generates interpretable feature representations, to extract domain-specific subsets from large datasets and apply structural pruning with domain-specific constraints. This ensures that the pruned model retains essential knowledge of the target domain. Furthermore, it leverages SAE-curated datasets to perform its own data distillation to recover key domain-specific information lost during the pruning process.

Takeaways, Limitations

FineScope outperforms several large-scale state-of-the-art LLMs on domain-specific tasks.
The pruned model recovers much of its original performance when fine-tuned on the SAE curated dataset.
Fine-tuning the pre-trained LLM without pruning the SAE curated dataset can improve domain-specific accuracy.
The Limitations of the paper is not specified.
👍