permissive-teachers.md

Permissive tabular teachers: TabICL and Mitra

Every teacher in this table may be distilled and the student shipped commercially: TabICL checkpoints are BSD-3-Clause, Mitra is Apache-2.0, TabPFN-2 is under the Prior Labs License (distillation permitted with attribution). Same datasets, splits, and caps as tabarena-full.md; xgboost/TabFM/TabPFN columns come from the other campaigns.

Device: mps. First calls include one-time weight download.

dataset	task	xgboost	TabPFN-2	TabICL	Mitra	gbt<-TabICL	soft<-TabICL	soft-oof<-TabICL	gbt<-Mitra	s/call (TabICL)
blood-transfusion	cls	0.711	0.765	0.775	0.775	0.765	0.759	0.759	0.765	1.6
diabetes	cls	0.740	0.750	0.760	0.750	0.766	0.740	0.760	0.729	1.03
anneal	cls	0.987	0.991	0.987	0.978	0.987	0.987	0.982	0.982	1.2
QSAR_fish_toxicity	reg	0.949	0.918	0.907	0.905	0.912	-	-	0.915	59.76
credit-g	cls	0.752	0.784	0.760	0.764	0.728	0.736	0.748	0.788	1.06
maternal_health_risk	cls	0.807	0.831	0.843	0.780	0.787	0.756	0.705	0.748	0.75
concrete_compressive_strength	reg	5.047	4.198	4.115	5.002	4.956	-	-	5.636	0.83
qsar-biodeg	cls	0.871	0.890	0.886	0.871	0.845	0.841	0.856	0.837	1.62
healthcare_insurance_expenses	reg	5217.371	4829.207	4458.266	4550.568	4458.082	-	-	4530.398	0.93
website_phishing	cls	0.885	0.909	0.920	0.888	0.894	0.888	0.879	0.876	1.01
Fitness_Club	cls	0.747	0.781	0.784	0.781	0.781	0.781	0.784	0.787	0.91
airfoil_self_noise	reg	1.613	1.092	1.027	1.653	2.229	-	-	2.318	0.99
used-Fiat-500	reg	830.136	803.507	784.192	841.820	794.371	-	-	835.630	1.1
MIC	cls	0.848	0.867	0.861	0.856	0.843	0.851	0.859	0.853	1.89
Is-this-a-good-customer	cls	0.880	0.893	0.893	0.893	0.891	0.888	0.891	0.893	1.12
Marketing_Campaign	cls	0.869	0.891	0.893	0.893	0.864	0.891	0.885	0.880	1.26
hazelnut-contaminant	cls	0.901	0.957	0.965	0.928	0.891	0.883	0.899	0.885	1.7
seismic-bumps	cls	0.931	0.936	0.936	0.936	-	0.936	0.936	-	1.1
splice	cls	0.965	0.965	0.976	0.963	0.955	0.955	0.957	0.952	1.79
Bioresponse	cls	0.781	0.773	0.792	0.795	0.768	0.787	0.776	0.795	1.94
hiva_agnostic	cls	0.968	0.968	0.968	0.968	-	0.968	0.968	-	1.9
students_dropout	cls	0.728	0.765	0.757	0.752	0.749	0.747	0.749	0.760	1.7
churn	cls	0.933	0.944	0.952	0.933	0.944	0.944	0.936	0.928	1.2
QSAR-TID-11	reg	1.509	1.413	1.393	1.526	1.445	-	-	1.527	1.94
polish_bankruptcy	cls	0.944	0.939	0.939	0.944	0.944	0.939	0.944	0.944	1.69
wine_quality	reg	0.783	0.654	0.634	0.669	0.666	-	-	0.675	1.1
taiwanese_bankruptcy	cls	0.968	0.976	0.979	0.979	0.971	0.971	0.971	0.973	1.85
NATICUSdroid	cls	0.880	0.899	0.904	0.885	0.896	0.893	0.880	0.893	1.91
coil2000_insurance	cls	0.944	0.952	0.952	0.952	-	0.952	0.952	-	1.89
Bank_Customer_Churn	cls	0.837	0.875	0.872	0.877	0.869	0.875	0.877	0.867	1.32
heloc	cls	0.757	0.776	0.773	0.765	0.771	0.752	0.757	0.760	1.28
jm1	cls	0.776	0.805	0.805	0.797	0.795	0.800	0.795	0.797	1.4
E-CommerceShipping	cls	0.645	0.693	0.693	0.707	0.685	0.691	0.699	0.707	0.95
online_shoppers_intention	cls	0.883	0.901	0.901	0.891	0.893	0.901	0.891	0.893	1.07
in_vehicle_coupon	cls	0.632	0.643	0.669	0.683	0.629	0.643	0.648	0.667	1.28
miami_housing	reg	149102.718	83065.624	90764.039	96749.143	195469.481	-	-	105196.030	1.17
HR_job_change	cls	0.728	0.744	0.731	0.747	0.755	0.733	0.741	0.741	0.89
houses	reg	0.268	0.239	0.226	0.249	0.261	-	-	0.268	0.88
superconductivity	reg	14.182	11.978	11.852	15.355	15.483	-	-	16.155	1.92
credit_card_default	cls	0.800	0.827	0.835	0.829	0.832	0.832	0.829	0.824	1.59
Amazon_employee_access	cls	0.928	0.939	0.939	0.939	-	0.939	0.939	-	0.96
bank-marketing	cls	0.856	0.867	0.867	0.875	0.867	0.867	0.869	0.877	0.91
Food_Delivery_Time	reg	8.720	7.445	7.493	7.514	7.458	-	-	7.505	0.85
physiochemical_protein	reg	5.071	4.491	4.414	4.814	5.059	-	-	5.051	0.78
kddcup09_appetency	cls	0.981	0.981	0.981	0.981	-	0.981	0.981	-	1.89
diamonds	reg	878.255	707.699	669.062	897.465	928.764	-	-	957.403	1.03
Diabetes130US	cls	0.904	0.917	0.917	0.917	0.917	0.917	0.917	-	1.84
APSFailure	cls	0.995	0.995	0.992	0.992	0.997	1.000	0.997	0.992	2.54
SDSS17	cls	0.952	0.960	0.957	0.965	0.960	0.963	0.968	0.960	1.74
airline_satisfaction	cls	0.909	0.920	0.909	0.912	0.885	0.891	0.883	0.875	1.74
GiveMeSomeCredit	cls	0.931	0.939	0.936	0.936	0.933	0.939	0.931	0.936	1.23

Win counts

• TabICL beats xgboost: 45/51
• TabICL beats TabPFN-2: 25/51
• Mitra beats xgboost: 37/51
• our gbt<-TabICL beats xgboost: 30/46
• out-of-fold soft labels beat in-context soft labels: 17/38

The soft-oof column uses stratified out-of-fold teacher labels, the fix from Tanna et al., "Pocket Foundation Models" (arXiv:2605.18654), for in-context teachers whose soft targets collapse toward one-hot on rows already in their context. On this suite it does not lift accuracy (win count above; median delta 0.000), and we measured why: the collapse is teacher-dependent, and TabICL v2 barely leaks here. Its in-context labels average 0.899 max-probability and 0.249 entropy (collapse would be near 1.0 and near 0), and they agree with the training labels at 92.5% against an 87.5% held-out accuracy: about five points of leak, where the paper's teachers memorize their context nearly perfectly. MLP students show the same null as gbt students, consistent with mild leak rather than student insensitivity. The practical diagnostic: compare your teacher's agreement with its own training labels to its held-out accuracy; that gap is the leak, and out-of-fold labeling is worth its 5x teacher cost when the gap is large.