Folder /data

I collect several famous russian classic novels (War and peace, Anna Karenina, Crime and punishment etc) in txt and stacked them together.

The corpora is approximetely 18MB.

Folder /models

I tried the most popular RNN architectures: GRU and LSTM with different capacity.

I made simple laguage models, so the quality metric is perplexity.

Dropout have improved LM significantly. However, LSTM + dropout performs almost the same as GRU + dropout. Thus, use dropout to achieve better performance, epspecially on a small corpora.