OSCE Minsk Group Statements Bot
here and follow the bot on Twitter.Read the statements
Statements are generated using the awesome char-rnn by Andrej Karpathy. We trained Justin Johnson's torch-rnn implementation on the original OSCE Minsk Group statements (~200KB English text) using the following parameters:
th train.lua -rnn_size 512 -dropout 0.5 -input_h5 ../bot/minskgroup/20160406.h5 -input_json ../bot/minskgroup/20160406.json
Statements are generated at quite low temperature (probably because of the small training set it doesn't learn very well):
th sample.lua -length 4000 -temperature 0.5 -start_text "VIENNA, 7 April 2016" -checkpoint ../bot/minskgroup/cv/s512d0.5_3400.t7
The model (25MB) is available in the repository. Hmm.. this is a quite heavy bot.