PEFT of SLM for Telecommunications Customer Support: A Comparative Study of LoRA Configurations with Energy Consumption Analysis

In three linesSystematic study of parameter-efficient fine-tuning (LoRA) on Qwen2.5-3B for telecom customer support. 16 LoRA configurations evaluated on 30,000 synthetic examples generated via Gemini 2.0 Flash. Key finding: validation loss (0.5024) does not correlate with quality ranked by GPT-5.2 and Claude 4.5 Sonnet judges. Energy consumption analysis included.

Summary generated by Claude — human-verified