perf: subsample to 1000 rows on HF and fix encoder consistency for correct predictions

Initial deployment of ModelMatrix-HF