How many churches are in the USA?
Nobody really knows for sure!
The IRS Exempt Organziations Business Master File contains around 270,000 "church-like" entities. A lot of non-church ministries and houses of faiths for entirely different religions get mixed in due to poor data entry practices at the IRS. A lot of churches inherit their denominations EIN, and many others never actually file at all (as they're not legally required to)
The US Religion Census says 350,000 churches, but they hold their data tightly so it's not independently verifiable.
The FEMA US Structures map says there are around 400,000 church buildings, but their data includes churches that have long been closed.
Overture Maps has 550,000 churches in their universe. These are largely sourced from Facebook pages because Meta donates data. But anybody can create a church page on Facebook and consequently have it end up in here.
I'm willing to average out the different counts at 375,000.
The bible says "where two or more are gathered", but for the purposes of this project, a church is defined as an entity where a regular gathering of christians occurs and collected funds ensure the continuation of the congregation. Building ownership isn't considered a filter, nor is absence from the IRS data (future versions of this project will handle churches from beyond the IRS universe)
Do big things with little LLMs
A tiny AI model (Qwen2.5~14b) running on a stock 16GB M4 Mac Mini was the engine behind getting this project started. Over 1 million prompts were executed through ollama to match church-like entities from the IRS data to their websites from purchased search results. An older, low-parameter model can't achieve 100% perfection, especially with the wide variation in church website quality, but the original relevancy prompt delivered tens of thousands more church website matches then were originally predicted.
It won't be possible to run enrichment and sermon scoring on my own hardware (especially given the memory shortages), but if I could get my hands on a 96gb machine, a higher parameter Qwen would be my first choice.