Official AAPA release: processed training data and A-GRPO checkpoints for adversarially anchored preference alignment.
Note 8B A-GRPO checkpoint.
Note 0.6B A-GRPO checkpoint.