I forgot to consider that I was turning this code over when I wrote it, so to help make sense of the bloated mess here are some notes:

If you want to run it:
1. Download a wikipedia XML, I recommend: http://mattmahoney.net/dc/enwik9.zip due to the size, saved as wikidata/enwik9
2. Run wikidata/process_data.py, it should fill the wikidata folder with 120 data_{n}.txt files
3. Run main.py
4. Optional: Run train_more.py
5. Adjust and run predict.py at you leisure


Some notes about most files in case they cause headscratching:

main.py
Run once to generate a model, do some initial training and output

train_more.py
Run repeatedly after main.py to do more intense training

predict.py
Uses the saved one_step_model (from main or train_more) to generate text.

play_maker.py
If the code has been adjusted to use the shakespeare data, this makes plays, otherwise mostly nonsense.

parameter.py
An attempt to isolate the important parameters for some exploration.

data/data.py
A databundle as I wanted to simplify passing data things around. (needs to be rethought)

my_model/construct.py
Picks a model and constructs it, this is to have one piece of code in charge of what model everything uses.

my_model/model_two_lstm.py
The model used for most testing. The others can be ignored.

one_step/one_step.py
Model for compounding predictions.