Update architecture_records/004-datapreprocessor.md

Co-authored-by: Will Johnson <[email protected]>
Signed-off-by: Dushyant Behl <[email protected]>

Author: dushyantbehl

architecture_records/004-datapreprocessor.md

@@ -220,7 +220,7 @@ Since this design allows complex preprocessing of the dataset on fly, the design
 
 The goal that we have here is to not be slower than the HuggingFace library which our whole design is based upon, in this sense we also imagine any performance improvements that we come across to be contributed back to HF library to keep our design simple and not reimplement stuff.
 
--> Handling Large Dataset
+#### Handling Large Dataset
 
 Our main reason for using HF [Map](https://huggingface.co/docs/datasets/en/process#map) heavily for data preprocessing is that for large datasets which are generally loaded as `IterableDatasets` the MAP API automatically performs [`lazy map operations`](https://huggingface.co/docs/datasets/en/about_mapstyle_vs_iterable#eager-data-processing-and-lazy-data-processing) and hence doesn't produce too much overhead while training.