Scholarlink

<p>CARD (Coding Assistant for Relational Data analysis) generates Python code that processes relational queries on raw data. Users can customize generated code via natural language instructions, e.g., by instructing the system to use specific libraries or produce certain output. Internally, CARD uses large language models such as GPT-4o to synthesize code. CARD automatically constructs prompts describing code generation tasks to the language models. Those prompts contain information on data format, customization requirements, as well as processing plans, generated by CARD’s scenario-specific query planner. CARD automatically tests generated code by comparing its output to the output of a reference SQL engine. In case of inconsistencies, CARD re-generates code with a certain degree of randomization. Furthermore, CARD can automatically generate libraries of code samples for specific customization scenarios in a pre-processing step, leveraging those samples at run time for few-shot learning. The experiments show that CARD generates accurate code in the vast majority of scenarios. Furthermore, current trends in language models are likely to benefit CARD’s performance in the future.</p>

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

Generating highly customizable python code for data processing with large language models

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Immanuel Trummer

摘要

CARD (Coding Assistant for Relational Data analysis) generates Python code that processes relational queries on raw data. Users can customize generated code via natural language instructions, e.g., by instructing the system to use specific libraries or produce certain output. Internally, CARD uses large language models such as GPT-4o to synthesize code. CARD automatically constructs prompts describing code generation tasks to the language models. Those prompts contain information on data format, customization requirements, as well as processing plans, generated by CARD’s scenario-specific query planner. CARD automatically tests generated code by comparing its output to the output of a reference SQL engine. In case of inconsistencies, CARD re-generates code with a certain degree of randomization. Furthermore, CARD can automatically generate libraries of code samples for specific customization scenarios in a pre-processing step, leveraging those samples at run time for few-shot learning. The experiments show that CARD generates accurate code in the vast majority of scenarios. Furthermore, current trends in language models are likely to benefit CARD’s performance in the future.