What's New
PPO dataset parameter fix for trl >=0.28
In newer versions of trl (>=0.28), PPOTrainer.__init__() no longer accepts dataset= as a keyword argument. This caused a TypeError when running soup train with task: ppo.
Fix: Now uses inspect.signature() to check whether the constructor accepts train_dataset, dataset, or neither. If neither is accepted, the dataset is set directly on the trainer object before calling .train().
Tests: 4 new tests in test_bugfixes.py (643 total).
Install / Upgrade
pip install --upgrade soup-cli🤖 Generated with Claude Code