ChatArena (or Chat Arena) is a Multi-Agent Language Game Environments for LLMs. The goal is to develop communication and collaboration capabilities of AIs.
Apache License 2.0
1.33k
stars
129
forks
source link
Suppress`transformers` "Models won't be available" message, update Umshini environments #87
This message previously printed every single time you imported chatarena, because the init.py imported the hf backend, which in turn tried to import transformers. I have now suppressed that error message, as people who are using that backend will be able to tell that they do not have pytorch or tensorflow when they try to use a given model.
For the umshini environments I have updated all of the prompts for the moderator judging, added info about max characters per response in the opening message (4000 char by default), and added print statements VIOLATION: True. so it is more transparent how these environments are judged. I tested many different prompts for the debate environment judge, and have it in a place where it does successfully penalize very short responses, responses where the agent simply agrees with the opponent, or cases where it tries to talk to the moderator directly.
This message previously printed every single time you imported chatarena, because the init.py imported the hf backend, which in turn tried to import transformers. I have now suppressed that error message, as people who are using that backend will be able to tell that they do not have pytorch or tensorflow when they try to use a given model.
For the umshini environments I have updated all of the prompts for the moderator judging, added info about max characters per response in the opening message (4000 char by default), and added print statements
VIOLATION: True.
so it is more transparent how these environments are judged. I tested many different prompts for the debate environment judge, and have it in a place where it does successfully penalize very short responses, responses where the agent simply agrees with the opponent, or cases where it tries to talk to the moderator directly.