Datasetpublic

Cameron-Chen/polaris-53k

Polaris 53K: conflict-free RL and GEPA splits This dataset derives deterministic, mutually disjoint splits from kushasareen/sdpo-datasets/polaris_5k at source commit 51fd81230be64a06d77e7edc33d3c0d60b3186b5. Splits Split Rows Intended use train (train_split.parquet) 51,911 RL actor training and GEPA trajectory training validation (val_split.parquet) 100 Fixed held-out GEPA prompt evaluation test (test.parquet) 500 Final model evaluation… See the full description on the dataset page: https://huggingface.co/datasets/Cameron-Chen/polaris-53k.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes40downloads

Dataset Card

Polaris 53K: conflict-free RL and GEPA splits

This dataset derives deterministic, mutually disjoint splits from `kushasareen/sdpo-datasets/polaris_5k` at source commit 51fd81230be64a06d77e7edc33d3c0d60b3186b5.

Splits

Split	Rows	Intended use
`train` (`train_split.parquet`)	51,911	RL actor training and GEPA trajectory training
`validation` (`val_split.parquet`)	100	Fixed held-out GEPA prompt evaluation
`test` (`test.parquet`)	500	Final model evaluation

Validation is sampled from the cleaned source train data. Test is sampled from the cleaned source test data. Both use seed 42.

De-duplication and conflict policy

The normalized problem key applies Unicode NFKC normalization, normalizes line endings, and collapses whitespace while preserving case.

—If one normalized problem has multiple normalized ground-truth answers, every occurrence of that problem is removed globally.
—If repeated normalized problems have the same normalized answer, the first occurrence is retained.
—If a consistent problem occurs in both source train and source test, the test occurrence is preserved and the train occurrence is removed.
—The three published splits have no normalized-problem overlap.

The build removed 26 train rows belonging to 13 conflicting problem groups, 235 additional duplicate train rows, and 6 train rows overlapping source test. No duplicate or conflicting rows were found within the source test data.

See split_manifest.json for source and output SHA-256 hashes, exact audit counts, normalization details, and split-key hashes.

Schema and provenance

The source parquet schema is preserved, including prompt, reward_model, and extra_info. Only extra_info.split is updated to identify each output split. The source dataset is derived from `POLARIS-Project/Polaris-Dataset-53K`.