Matches the training dtype and the JAX release; the previous float32 tensors were an inflated cast of the same bf16 values (roundtrip-exact), so this halves the download losslessly.
Converted from google/tabfm-1.1.0-jax via the official convert_and_upload pipeline (per-parameter mapping, strict load, parity-gated: regression 1.6e-06, classification 7.0e-04 f32 dummy-input parity, the latter fully attributed to per-op precision-pinning differences compounding over the 24 ICL blocks; clean-float64 identical-input divergence ~1e-5).
