Four-month LiveMath package built from the v5 -> ge5 -> hard pipeline.
Months included:
202511: full=337, ge5=141, hard=29, hard_acc=0.4138
202512: full=472, ge5=213, hard=52, hard_acc=0.2885
202601: full=403, ge5=192, hard=46, hard_acc=0.3913
202602: full=390, ge5=188, hard=50, hard_acc=0.4000
Overall hard-set accuracy on gpt-5.4_2026-03-05 medium:
65/177 = 0.3672
Per-month layout:
full/qaEval_
full.json
ge5/qaEval_ge5.json… See the full description on the dataset page: https://huggingface.co/datasets/hendrydong/livemath-v7-0316. 
