Creating specialized large language models (LLMs) demands massive amounts of clean, tailored data. However, filtering web-scale datasets with powerful LLMs like GPT-4 is incredibly expensive. Researchers have introduced SIEVE, a new system that mimics GPT-4's filtering accuracy at a fraction of the cost. SIEVE combines a lightweight T5 model with a clever active learning algorithm. This algorithm strategically selects the most informative snippets from the massive dataset and queries GPT-4 only on those, drastically cutting down on the expensive API calls. After this targeted training, the T5 model becomes a cost-effective filter, achieving near-GPT-4 performance at a tiny fraction of the computational expense. Experiments on the OpenWebText dataset, using various custom filters for topics like politics, climate change, and AI, confirmed that SIEVE achieves comparable or even better filtering results than GPT-4. Human evaluators actually preferred SIEVE's filtering in some cases. This breakthrough has significant implications for developing specialized LLMs. It makes curating high-quality datasets far more affordable, potentially opening doors to faster innovation in tailored AI models for specific industries and applications. Future research will focus on scaling SIEVE to handle even larger datasets and other data types like images and audio, making SIEVE a key tool in the responsible development of AI.