Refining Large Language Models for Tabular Data Analysis in Business Domain by Laymen Text

Organizations face various challenges when analyzing tabular data. One of the key challenges is the complexity of business data analytics tasks. These tasks involve handling large volumes of data, organizing and structuring it, and extracting valuable insights. Additionally, employees who work in financial analysis often encounter a significant workload on tools such as Excel, SAP, PowerBI, and Tabula. This workload can result in increased effort and time required to analyze and make sense of the data. Organizations must address these challenges to ensure efficient and effective analysis of tabular data. Organizations spend more money to perform business tasks. Nowadays, there are many efficient models in artificial intelligence to perform text, audio, video, and image-based tasks, but no efficient models are available to perform tabular-based tasks specifically. Pandas Python library provides various functionalities and APIs that are useful for business data analysis. This research solved the problem with tabular data using Python Pandas code generation. Here, the researchers used two datasets, each with 50 records. The Large Language Model (LLM) is a supervised Learning Pre-Trained Foundation Model (SSL PFM) category based on text generation. The SSL PFM helps the language models learn the context of language and world knowledge. During this research, models such as LLaMA-2, Falcon, CodeLlama, and Mistral were considered for analysis. Each of these models consists of a Billion parameters. Moreover, Quantization techniques were incorporated to reduce model size, enabling models to load with minimal hardware. After quantizing the model, Parameter Efficient Fine Tuning (PEFT) trains the dataset using only a few model layers; other layers are frozen. Well-known experts with Pandas evaluate the fine-tuned models of chosen language models. Finally, mistral-7B produced prominent results in analyzing business tasks and producing summarized results.