2506.10885

Total: 1

#1 Slimming Down LLMs Without Losing Their Minds [PDF1] [Copy] [Kimi2] [REL]

Authors: Qingda, Mai

This paper investigates and validates the impact of fine-tuning on large language model performance, focusing on parameter-efficient methods (LoRA and QLoRA). We evaluate model capabilities across three key domains: (1) commonsense reasoning (HellaSwag), (2) mathematical reasoning (GSM8K), and (3) multi-domain knowledge (MMLU-CS). Our findings demonstrate that: (1) LoRA-based methods effectively improve task-specific performance while maintaining computational efficiency, and (2) performance strongly depends on alignment between fine-tuning dataset and benchmark tasks. The study provides both theoretical insights into parameter-efficient mechanisms and practical guidance for developers implementing efficient LLM adaptation with limited resources.

Subjects: Computation and Language , Artificial Intelligence

Publish: 2025-06-12 16:49:40 UTC