Datasetpublic

jucamohedano/oxford-pets-grpo-think

Oxford-IIIT Pet — GRPO training data (structured reasoning) GRPO (verl) training data for Oxford-IIIT Pet breed classification with a structured-reasoning prompt: the model emits a scratchpad tagging visible properties (HasProperty), parts (HasA), and setting (AtLocation) before the label. Reward: 0.30 for a well-formed think block, 0.70 for the label match. Splits: train 2,944 rows, test 3,669 rows. Schema column type data_source string prompt… See the full description on the dataset page: https://huggingface.co/datasets/jucamohedano/oxford-pets-grpo-think.

sourceHugging Facecc-by-sa-4.0updated 2mo agoView on Hugging Face

0likes31downloads

Dataset Card

Oxford-IIIT Pet — GRPO training data (structured reasoning)

GRPO (verl) training data for Oxford-IIIT Pet breed classification with a structured-reasoning prompt: the model emits a <think> scratchpad tagging visible properties (HasProperty), parts (HasA), and setting (AtLocation) before the label. Reward: 0.30 for a well-formed think block, 0.70 for the label match.

Splits: train 2,944 rows, test 3,669 rows.

Schema

column	type
`data_source`	string
`prompt`	list<element: struct<content: string, role: string>>
`ability`	string
`reward_model`	struct<ground_truth: string, style: string>
`images`	list<element: struct<bytes: binary, path: string>>
`extra_info`	struct<gt_label: string, index: int64, split: string>

Format

Parquet in the verl RL schema, one row per image: a chat prompt (system + user), an images column (PNG-encoded), reward_model = {style: rule, ground_truth: <breed>}, ability: classification, and extra_info. Consumed directly by the verl GRPO trainer with data.image_key=images.

Provenance

Built with generate_grpo_dataset() for GRPO training experiments on open-world image classification with large multimodal models (Test-Time Warm-Up line). Derived from the Oxford-IIIT Pet Dataset (Parkhi et al., CVPR 2012).