Add finetuning data

[codethazine]

• Augmented the data 3x with the OAI API on gpt-3.5-turbo with the following prompt: Transform this viral tweet into {opposite} version: '{tweet}'" where opposite is one of the following: ["an opposite sentiment", "a boring", "a short and non-viral"]
• Fixed data dir structure
• Added finetuning data on model/data/final

[codethazine]

Finetuned with the data on OAI Curie and ran some tests. Better than the OAI prompted server, but deff still needs some fiddling with:

Extra tests:

1. Add extra data, currently fine-tuned on ~500 tweets, e.g. getting it to ~1000
2. Change data augmentation logic, e.g. extract what tweet is about and use that as training prompt
3. Change base model, e.g. curie -> davinci, or llama2