I am currently building my own dataset for the go bot and i was wondering how many conversations are required to get decent results ? i’ve seen in the tutorial that you used only 50 dialogs for training but no mention for test and validation, i’m willing to do 100 convs (60/20/20) but i’m not sure if this will be enough
Hey averroes,
More than the number of dialogues, the length per dialogue i.e. how often a slot & go bot response appears is critical for performance.
Based on my experience you should aim for all slots appearing >9 times and >4 examples of each possible response uttered by the go bot across all dialogues. E.g. if you have the slots “food”, and “telephone”, just make sure each of them is used like 10 times. And if you have the responses “Hello” and “Thanks”, make sure they both appear like 5 times.
Hope this helps =)
1 Like
Thanks a lot, i’ll update you with the result i got, i’m aiming for go_bot_dstc_best configs, i saw that it uses fastext, does it give better result or should i just give up on it since it takes so much RAM ?