DistillBART Metrics

	B	C	D	E	F	G	H	J
1
2		MM Params	Inference Time (MS)	Speedup	Rouge 2	Rouge-L
3	distilbart-xsum-12-1	222	90	2.54	17.98	33.31
4	distilbart-xsum-6-6	230	132	1.73	21.17	36.21	Update
5	distilbart-xsum-12-3	255	106	2.16	22.40	37.30	*Trained with pegasus pseudo-labels
6	distilbart-xsum-9-6	268	136	1.68	22.08	37.24
7	bart-large-xsum (baseline, 12-12)	406	229	1.00	22.29	37.20		0.1823
8	distilbart-xsum-12-6	306	137	1.68	22.32	37.39
9
10	bart-large-cnn (baseline, 12-12)	406	381	1.00	21.06	30.63
11	distilbart-12-6-cnn	306	307	1.24	21.26	30.59
12	distilbart-6-6-cnn	230	182	2.09	20.17	29.70
13	distilbart-12-3-cnn	255	214	1.78	20.57	30.00
14
15	pegasus-cnn_dailymail (16-16)		1400		21.37	30.94
16	dpx-16-4-cnn		436	3.21	21.29	31.3
17
18	pegasus-xsum (16-16)		373		24.46	39.1507
19	dpx-16-8-xsum		195	1.91	23.25	38.03
20	distill-pegasus-xsum-16-4				23.18
21
22	Inference Time is Samples/Second in fp16 on a v-100 GPU with bs=16. For CNN I only timed 1000 samples. Pegasus is run in fp32, it breaks otherwise
23	ROUGE is measured using pyrouge, and differs slightly from the original bart paper
24	Code is here
25	The speedups are not identical between datasets because some models generate longer summaries than others.
26	More granular timing info for CPU/GPU: here
27	*Update 2020-09-03: updated xsum numbers (higher across the board) after #6526
28	Paper: https://arxiv.org/abs/2010.13002
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100