TRL version / dependency mismatch when running train_grpo.py

[Clalloures]

I’m currently trying to run the BioReason codebase locally and ran into some issues when executing the training script train_grpo.py. I was able to install the dependencies listed in the repository, but the script fails at runtime.

From my debugging so far, it looks like the issue might be related to TRL version incompatibilities, possibly due to API changes across different TRL releases.

Questions

I wanted to ask for clarification on a few points:

Which TRL version was used (or is recommended) to run BioReason correctly?

Are there any specific dependency versions (Transformers, Accelerate, PyTorch, etc.) that are particularly important to pin?

From a learning/debugging perspective, is there a recommended order to study or run the .py files (e.g., dataset processing → model definition → training scripts)?

Context

Script: train_grpo.py

Environment: local setup (Python + venv)

Issue appears even after installing dependencies, likely due to TRL API differences

I’m happy to provide logs, stack traces, or environment details if that would help.

Thanks again for making the code public and for any guidance you can share!

[bminesh-shah]

How were you able to procure the base model weights? @Clalloures

[DBpackage]

I'm also undergoing this issue.

The authors should have provided the specific versions for the environment setup.