Table 4:

# of candidates pooled for each training instance. $m$ is % of noun phrases masked, $s$ % of entities swapped, and $t$ the softmax temperature for GPT-3.

Method	Hyper-Param	Number
Mask-And-Fill (Low)	m = 0.25	10
Mask-And-Fill (High)	m = 0.75	10
Swap Intrinsic (Low)	s = 0.5	10
Swap Intrinsic (High)	s = 1.0	10
Swap Extrinsic (Low)	s = 0.5	10
Swap Extrinsic (High)	s = 1.0	10
Paraphrase	t = 0.7	5
Reference	N/A	1
Total For Faithfulness		66
Diverse Beam (PRIMERA)	p = 1	10
Diverse Beam (LongT5)	p = 1	10
Total For Relevance		20