LSTM autoencoder input specification for reconstruction task


I got a question while implementing LSTM auto-encoder.
I have data that with 10000 rows and 60 columns in a one-hot encoded form.
How can I feed it to the LSTM encoder such that from the latent space encodings decoder can reconstruct the entries?

PS.- I want to feed 4 entries together to the encoder every time the time-stamp is increased.

Any help would be appreciated!