Issues with different versions of GPT

[martin-krutsky]

In reverie/backend_server/persona/prompt_template/run_gpt_prompt.py, multiple requests to OpenAI are made with a hardcoded model gpt-35-turbo-0125, which is currently not a valid/supported model on the OpenAI API (unlike on Azure).
If I try to run the simulation (e.g., with the command run 1) with the name gpt-3.5-turbo-0125, the simulation seems to be working.

However, if I change the model to a newer (and cheaper) one, e.g., gpt-4o-mini-2024-07-18, I eventually get the following error:

  File "...generative_agents\reverie\backend_server\persona\memory_structures\spatial_memory.py", line 107, in get_str_accessible_arena_game_objects
    x = ", ".join(list(self.tree[curr_world][curr_sector][curr_arena.lower()]))
KeyError: '{main room'

caused by the output of the gpt call being {main room without a closing bracket, i.e. wrong syntax on the model's output.

Do you know more about which models have the prompts been successfully tested on (e.g., the original joonspk-research's repo contained yet another model, text-davinci-002), what was your reasoning behind choosing gpt-3.5-turbo-0125, and do you have any suggestion for resolution of the problem for other models (e.g., structured output comes to mind)?

[drudilorenzo]

Unfortunately I have not tested models other than gpt-3.5-turbo-0125.

The reason why I choosed that model is that text-davinci-002 was already deprecated, and gpt-3.5-turbo-0125 was the cheapest model available at that time.

I suggest you to look to the changes made by @chowington in his pull request #8. You may find the solution there.