Note The training dataset used for the GRPO and RAFT models

Note A companion dataset to Aletheia-Train containing preference pairs for contrastive learning

Note A completely in-distribution test set for our models

Note An out-of-distribution test set with codes generated by stronger models

Note An out-of-distribution test set with more difficult code comparisons

Note An out-of-distribution test set modified to exploit LLM biases and test verifier robustness