HOW-TO.md

How To

Steps in running the code

The model is trained using the privacyFineTune.ipynb file and is queried with the testPrivacy.ipynb file. There is no need to run the fine-tuning again as the model is already avalible on huggingface however we will still give instruction on how to do it.

fine-tuning

Fine tuning is pretty simple as the majority of it has alread been set up for easy use. If you wish to train the model the same as we did, just run everything as is and then log in using your personal huggingface token.

If you wish to use a different model or dataset, then do the following

• As this file is made specifically for fine-tuning Llama 2, we suggest keeping the model_name variable untouced. However if you want to change it, change it to the model name you wish avalible in the Transformers Library.
• Change the dataset name to whatever dataset is desired (must be supported by huggingface dataset library).
• Change the new_model name to whatever you want.
• adjust the dataset_text_field="" variable to the name of the row of text in your dataset. For example in the sjsq dataset the text column is called "Text".
• Change the prompt variable to a question you want to ask it.

Aside from these changes, run the file from top to bottom to train the model. Should take about 25 minutes in total.

Prompting

The testPrivacy.ipynb file contains the test prompts that were used to test the model. It should be set to run as is with our model. If you wish to add custom prompts to the file, do so by creating a new codeblock and using the syntax

prompt = "Your question"

We also provided two different privacy policies for refrence. The first is from TopHive and the second is from Starbucks. The starbucks one does not work as it is too big and you run out of GPU ram fast on the free colab plus Llama doesn't like how many words are in it. TopHive does work however. To use these privacy policies in your prompt change the policy variable to the name of the company you are using the policy of

policy = starbucks

Then run one of the question boxes or make your own prompt.

To run the text generation chose a prompt first by running that box then run

result = pipe(f"<s>[INST] {prompt} [/INST]")
print('\n',result[0]['generated_text'])
resultList.append(result[0]['generated_text'])

Other than that basically run the code from start to bottom until you get to the prompt section. All prompts are saved in the "resultList" variable.