Using the cldr (and contributing to it, though that can be frustrating) is likely better than trying to reinvent its dataset. It has a number of issues which I

Using the cldr (and contributing to it, though that can be frustrating) is likely better than trying to reinvent its dataset. It has a number of issues which I don’t think it properly warns about (well they’re more LDML issues) e.g. the original date time pattern system had significant cross-cultural compatibility issues if you wanted something other than the main date/time patterns. The skeletons system improves that a lot, but is decidedly less complete. Also some placeholders are straight up unusable e.g. it was a month-week system which was designed off of the iso weeks system, but doesn’t provide a “month” placeholder for it, so W is literally just a trap.