Complete result settings for Reuters-52 dataset (table columns are sortable)
The experiments are described in Rooshenas and Lowd, Discriminative Structure Learning of Arithmetic Circuits, AIStats 16

Split Penalty	Standard Deviation	L1 Penalty	Initial Evidence	Train CLL	Validation CLL	Test CLL	Node#	Edge#	Feature#	Learning Time
2	0.5	2	10	-43.894033	-43.103935	-41.949274	27274	36028	9646	72157.060458
5	0.5	0.5	10	-43.853805	-43.130789	-41.978367	27294	36056	9654	72432.157637
2	0.5	0.5	10	-43.863670	-43.116286	-41.967610	27309	36077	9660	72843.163155
5	0.5	0.5	20	-42.787454	-42.618001	-41.543381	37999	51043	13936	75801.717386
2	0.5	1	10	-43.870311	-43.131539	-41.977188	27234	35972	9630	71499.309451
2	0.5	0.5	20	-42.787262	-42.616294	-41.544076	38004	51050	13938	75018.627434
10	0.5	0.5	10	-43.856424	-43.157338	-41.995267	27264	36014	9642	74247.319691
10	0.5	0.1	10	-43.843504	-43.144261	-41.978903	27274	36028	9646	73107.547963
5	0.5	1	10	-43.865155	-43.130321	-41.976790	27239	35979	9632	72299.083868
5	0.5	0.1	10	-43.849718	-43.139202	-41.986425	27304	36070	9658	71856.789107
10	0.5	0.1	20	-42.793693	-42.645138	-41.562268	37994	51036	13934	76854.196385
2	0.5	0.1	10	-43.857588	-43.144954	-41.993607	27314	36084	9662	73104.424438
10	0.5	1	10	-43.866088	-43.133707	-41.977230	27219	35951	9624	72100.313085