Scholarlink

This study presents a comprehensive evaluation of fine-tuning applied to Small Language Models (SLMs) for specialization in highly specific scientific domains, with a case study in ecological monitoring. A comparative analysis was conducted between the Llama 3.2-3B model, adapted using a proprietary corpus on native flora, and the baseline DeepSeek-R1-Distill-Qwen-1.5B model. Quality was measured using the perplexity metric, yielding substantial improvements: the fine-tuned model achieved values of 1.25 and 1.60 on the DPAV1.0 and DPAV2.0 datasets, compared to 49.48 and 108.95 for the baseline model, demonstrating orders-of-magnitude superior contextual understanding and terminological precision. The experimental architecture integrates environmental sensors, cloud infrastructure, a REST API, and a conversational agent based on the specialized SLM, capable of processing real-time data and generating actionable recommendations. Computational performance analysis showed that, with a moderate consumption of 3 GB VRAM and competitive inference times, the fine-tuned model is viable for deployment on mid-range hardware, enabling adoption in resource-constrained environments. The results confirm that SLM fine-tuning is not only technically feasible but also a highly efficient strategy for creating expert AI agents in niche domains, offering a practical and scalable alternative to massive LLMs. This work sets a precedent for the development of specialized, accurate, and low-cost AI systems with direct applications in environmental conservation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fine-Tuning of Small Language Models for Ecological Monitoring: A Comparative Analysis Based on Perplexity

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jhon Edilberto Monroy Barrios,
  • Enrique Miguel Paco Cusi,
  • Miguel Angel Portella Huincho,
  • Jeymi Melanie Valdivia Eguiluz,
  • Basurco Zapata Gerald Francis,
  • Fiorella Jara Rosales

摘要

This study presents a comprehensive evaluation of fine-tuning applied to Small Language Models (SLMs) for specialization in highly specific scientific domains, with a case study in ecological monitoring. A comparative analysis was conducted between the Llama 3.2-3B model, adapted using a proprietary corpus on native flora, and the baseline DeepSeek-R1-Distill-Qwen-1.5B model. Quality was measured using the perplexity metric, yielding substantial improvements: the fine-tuned model achieved values of 1.25 and 1.60 on the DPAV1.0 and DPAV2.0 datasets, compared to 49.48 and 108.95 for the baseline model, demonstrating orders-of-magnitude superior contextual understanding and terminological precision. The experimental architecture integrates environmental sensors, cloud infrastructure, a REST API, and a conversational agent based on the specialized SLM, capable of processing real-time data and generating actionable recommendations. Computational performance analysis showed that, with a moderate consumption of 3 GB VRAM and competitive inference times, the fine-tuned model is viable for deployment on mid-range hardware, enabling adoption in resource-constrained environments. The results confirm that SLM fine-tuning is not only technically feasible but also a highly efficient strategy for creating expert AI agents in niche domains, offering a practical and scalable alternative to massive LLMs. This work sets a precedent for the development of specialized, accurate, and low-cost AI systems with direct applications in environmental conservation.