CNCL-Penn-State/MuCE-ORACL-Gemma

MuCE for ORACL Gemma

The MuCE responses, human ratings and exact inputs used for ORACL Gemma 4 31B. Training, validation and the four evaluation sets are kept separate: 170,445 training responses, 21,485 validation responses, 21,429 test responses, 17,979 responses to held-out items, 12,144 responses to held-out tasks and 10,128 responses in held-out languages.

messages contains the complete instruction and participant response, in their original order. facsco_grm_scaled is the human rating on the 10–50 scale. Public row IDs replace source identifiers. Extract images.zip beside the data before scoring drawing responses; image paths in the messages are relative to that directory. The original texts, ratings, image bytes and split assignments are unchanged.

The original MuCE dataset is described in Creative Preference Optimization. This release preserves the inputs used in the Gemma study and does not include the additional cohorts used in the separate full-data training run.

from datasets import load_dataset
data = load_dataset("CNCL-Penn-State/MuCE-ORACL-Gemma")