Derived from AhNr/dr-kernel-RL.
Rows, labels, reward_model, and extra_info are identical; the only change is the
shared prompt preamble, edited to teach the Ascend NPU execution model so the policy can
act on msprof feedback.
What changed (uniform across all rows):
Fixed the few-shot example to the NPU idiom: grid = (min(triton.cdiv(N, BLOCK_SIZE), 48),),
an in-core grid-stride loop, BLOCK_SIZE=8192, and multibuffer=True
(was CUDA-style… See the full description on the dataset page:
https://huggingface.co/datasets/brucexi99/dr-kernel-RL-v4.
