- Run
cd workbench/
- Run
npm run task -- ../prompt_injection_uplift/ <TASK_NAME>
where <TASK_NAME> is one ofgpt3_easy
,gpt3_medium
,gpt3_hard
, andgpt3_very_hard
. These tasks are defined inprompt_injection_uplift/prompt_injection_uplift.py
- Copy the container id from the output of the above comand to pass to the agent.
- Run
npm run agent -- <CONTAINER_NAME> example-agents/langchain-react-baseline-with-write-file/ "pip install -q -r requirements.txt && env OPENAI_API_KEY='sk-proj-...' MAX_STEPS=<MAX_STEPS> python -u main.py"
Note that you also need to pass MAX_STEPS. Currently they are mapped inside prompt_injection_uplift/prompt_injection_uplift.py
as follows:
Variant | Num of Steps |
---|---|
Easy | 30 |
Medium | 50 |
Hard | 70 |
Very Hard | 100 |
MAX_STEPS depends on which variant of the task you are running the agent on. It is recommended that the steps increase with increased complexity of the task