But context modelling isn't dictionary compression. Parent is correct that dictionary compression itself (e.g. LZ4) isn't very useful for floating point.