Hi,
I am a PhD researcher working on machine learning and specifically deep learning.
I have been working on several deep learning projects (both research and freelance projects), using Tensorflow and Python.
Concerning your questions:
1. I have used RNN and LSTM models in many of my projects, for different problems, from sentiment analysis, document classification to language modelling and computer vision.
2. I would propose to use Python3 (>= 3.5)
3. The training set should be determined based on your task. I do not have enough information about your task to suggest a concrete dataset. In principle, a dataset, which is as close as possible to the domain of your task, should be used.
4. For generating new texts from a trained model, a greedy decoder or a beam search decoder (more advanced) can be used.
5. For any of my projects, the deliverables contain code files with comments and readme files (.txt or .doc) describing the project structure and code usage.
If you are interested, please send me a message and we can further discuss.