Skip to content

Latest commit

 

History

History
64 lines (38 loc) · 3.09 KB

README.md

File metadata and controls

64 lines (38 loc) · 3.09 KB

Context:

These scripts are intended for the creation a training dataset to fine-tune an LLM nad make it resistant to prompt injection attacks, as described on WithSecure labs research blog: This was the output of research described in this WithSecure Labs article: https://labs.withsecure.com/publications/llama3-prompt-injection-hardening..

For further details see this TogetherAi API documentation.

For examination of the model created from this experiment see this link to view the WithSecure Huggingface profile.

Training - Usage guide [Simple]:

For an easy use, formatted for TogetherAi Llama3.1-8b:

Populate your emails.jsonl file with email "text", related "instruction" and desired "output"

Populate your Breakouts.txt file with the desired prompt-injection breakout text, placement of the malicious prompt should be marked by [XXX].

Populate Prompts.txt with your desired malicious prompts, making sure these can easily be examined for 'canary' values.

Execute Run.py, to run the following scripts automatically.

Training - Usage guide [Custom]:

The purposes of each provided python script are as follows:

Combine.py - combine each possible combination of breakout and prompt injection input - Output.txt is created containing -ALL- of these combinations.

Compile.py - randomly compile each prompt created from combine.py into an email payload. - this script will randomly decide if an email is selected to have a payload added. - the payload is added either, in the middle, at the end or at the end following two line breaks; for greater variety of injections. - dataset.jsonl is created.

Prep.py - renames the section labells and shuffles the order of your dataset. - adds data tags and the context question to the dataset. - datasetReady.jsonl is created.

Format.py - adds the system prompt and formats the dataset to suit the Llama3.1-8b formatting reqauired for fine-tuning with togetherAi - FormattedFinal.jsonl is created.

Calc.py - to calculate the dataset size, number of samples and details. - ensure the size of your dataset is appropriate before beginning training.

Validate.py - to ensure correct Together pip formatting (pass = good)

Testing - Usage guide:

Ensure a portion of the training dataset created is removed from your file, this is necessary for testing purposes. Your testing dataset should be saved under TESTselection.jsonl.

Compare.py - to cycle through and compare outputs from your base and fine-tuned models. - Ensure canary_words is upto date, warning: this variable may contain profanity. - TestOutputs.jsonl is created.

Examine.py - to examine the contents of output.jsonl file via excel for data collection purposes.

Notes:

Together PiP package is required.

LangDetect PiP package is required.

Never hard-code your API keys.

Always check foreign scripts before running them.