The x86 build of ML.NET has some remaining baseline differences that need investigation and fixing.
From initial investigation, the remaining differences have fewer than 4 matching significant digits and this may be caused by known bugs in the double parsing code (which have already been fixed or are actively being fixed in .NET Core). Other differences are due to the implementation (between x86 and x64) of some floating-point Math functions that the tests depend on.