Add the Appendix E.2 / Table 12 SFT setup, 32 rollouts per item before and after training, the 100,000-permutation one-sided stratified randomization test, and the reported gains and p-values. Explain the evaluation use case and distinguish the single-item paper experiment from the full augmented release.
Import AIME24 (50), AIME25 (52), AIME26 (53), and AMC23 (80) from ASTRAL-Group/benchmark-difficulty-measure at 53cf20cdb98dc9c2a4f780e582f00c559fb9e165. Preserve the original JSON files byte-for-byte, add JSONL with string answers for reliable dataset loading, and configure four test subsets with provenance and usage documentation.
