Making a synthetic dataset for fine-tuning

I've been thinking about building a pipeline to generate reasoning training data for LLMs, but I want to avoid the common failure mode of synthetic data where you just generate the same template with different numbers. The rough idea: Generate an abstract reasoning task (logic, planning, graph prob

中文全文翻译正在补齐，可先查看下方英文原文。

I've been thinking about building a pipeline to generate reasoning training data for LLMs, but I want to avoid the common failure mode of synthetic data where you just generate the same template with different numbers. The rough idea: Generate an abstract reasoning task (logic, planning, graph prob