LoRA based instruction fine tuning of large language models for agronomic triple extraction and GraphRAG knowledge retrieval

Due to massive growth in the unstructured agricultural data generated from reports, research articles, news, farmers’ feedback, experts’ advice, etc, there is a need to utilize this information efficiently for knowledge representation and decision-making. There are multiple techniques where information extraction and managing it in a structured manner can be less challenging. LLMs are used widely for extracting meaningful information in the way the user instructs. More precisely, LLMs, when fine-tuned on domain-specific data, e.g, agricultural data, help to retrieve information promptly. We propose a framework where the Llama model is fine-tuned using the LoRA technique and an instruction dataset for extracting triplets from the unstructured agronomic data. Further, this Llama-LoRA-based model is used to generate knowledge, which will efficiently help to retrieve the query results for the user using GraphRAG. Domain-specific instructional fine-tuning has shown a significant rise in performance. ROUGE and BLEU precision results of the proposed framework have outperformed the response retrieval compared to other retrieval techniques. Proposed results of domain-specific instructional fine-tuned Llama-LoRA model have overcome the limitations of knowledge representation and unstructured data extraction through GraphRAG, resulting in semantic search with contextual retrieval.