List your company

Robust Classification on Civil Comments (test)

92.2Average AccuracyERM
46.64858.47470.382.126Sep 19, 2022Apr 13, 2023Nov 5, 2023May 29, 2024Dec 21, 2024Jul 15, 2025Feb 6, 2026

Evaluation results

MethodLinks
ERMGroup-oblivious setting=true2022.0992.256——
ERM2023.0492.256.5——
Focal LossGroup-oblivious setting=true2022.0991.260.1——
Sparse and Shared Feature Activations2023.0491.275.45——
JTTGroup Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0291.169.3——
JTTGroup Info (Train/Val)=✗/★2026.0291.169.3——
LEIAGroup Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0290.972.9——
LEIAGroup Info (Train/Val)=✗/★2026.0290.972.9——
JTTGroup-oblivious setting=true2022.0990.767.4——
UMIXGroup-oblivious setting=true2022.0990.670.1——
DRO2023.0490.269——
chi^2-DOROGroup-oblivious setting=true2022.0990.163.8——
CVaR-DOROGroup-oblivious setting=true2022.099064.1——
GICGroup Info (Train/Val)=✗/★, Group labels usage in training=not used2026.029072.5——
GICGroup Info (Train/Val)=✗/★2026.029072.5——
AFRGroup Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0289.868.7——
AFRGroup Info (Train/Val)=✗/★2026.0289.868.7——
chi^2-DROGroup-oblivious setting=true2022.0989.464.2——
LISAGroup Info (Train/Val)=✓/★, Group labels usage in training=used2026.0289.272.6——
CVaR-DROGroup-oblivious setting=true2022.0989.162.3——
ERMValidation selection=minimax2026.0189.12.20.869—
GALAXYinitial_labels=100, query_schedule=10 x 102025.1087.77———
DFRGroup Info (Train/Val)=✗/✓✓, Group labels usage in training=used2026.0287.270.1——
Confidenceinitial_labels=100, query_schedule=10 x 102025.1086.43———
ParBaLS-MAP EPIGinitial_labels=100, query_schedule=10 x 102025.1085.67———
ParBaLS EPIGinitial_labels=100, query_schedule=10 x 102025.1085.59———
IRM*Group Info (Train/Val)=✓/★, Group labels usage in training=used2026.0285.563.2——
ERM*Group Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0285.463.2——
ERMGroup Info (Train/Val)=✗/★2026.0285.463.2——
CVaRDRO*Group Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0283.568.7——
CVaRDROGroup Info (Train/Val)=✗/★2026.0283.568.7——
CVaRValidation selection=minimax2026.0183.5230.605—
ReWeightCRT*Group Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0283.468.2——
ReWeightCRTGroup Info (Train/Val)=✗/★2026.0283.468.2——
CRT*Group Info (Train/Val)=✗/★, Group labels usage in training=not used2026.028368.5——
CRTGroup Info (Train/Val)=✗/★2026.028368.5——
LV-GroupValidation selection=minimax2026.0182.851.60.312—
CnCGroup Info (Train/Val)=✗/★, Group labels usage in training=not used2026.0281.768.9——
CnCGroup Info (Train/Val)=✗/★2026.0281.768.9——
OTTERBackbone=BERT, Setting=Zero-shot2024.0481.4———
Group DROGroup Info (Train/Val)=✓/★, Group labels usage in training=used2026.0281.477.7——
Randominitial_labels=100, query_schedule=10 x 102025.1080.88———
GLISTERinitial_labels=100, query_schedule=10 x 102025.1080.19———
χ2-DROValidation selection=minimax2026.0179.824.70.551—
BADGEinitial_labels=100, query_schedule=10 x 102025.1079.76———
GroupDROValidation selection=minimax2026.017745.60.314—
GroupDRO-BValidation selection=minimax, Training data=bulk-filtered2026.0176.845.30.315—
CoreSetinitial_labels=100, query_schedule=10 x 102025.1072.33———
DPEGroup Info (Train/Val)=✗/✓, Group labels usage in training=used2026.0271.582.2——
Prior MatchingBackbone=BERT, Setting=Zero-shot2024.0457.2———
Zero-shotBackbone=BERT, Setting=Zero-shot2024.0448.4———
CAMLAcquisition Function=Random, Backbone=BERT2026.05———49.2
CAMLAcquisition Function=Confidence, Backbone=BERT2026.05———58
CAMLAcquisition Function=CoreSet, Backbone=BERT2026.05———63.1
CAMLAcquisition Function=BADGE, Backbone=BERT2026.05———62.7
CB last-layer retrainingAnnotations Group=false, Annotations Class=true2023.09—80.4——
Class-balanced ERMAnnotations Group=false, Annotations Class=false2023.09—61.4——
DFRAnnotations Group=true, Annotations Class=true2023.09—78.8——
DFR (our impl.)Annotations Group=true, Annotations Class=true2023.09—81.8——
ES disagreement SELFAnnotations Group=false, Annotations Class=false2023.09—79.1——
Group DRO-ESAnnotations Group=true, Annotations Class=true2023.09—80.4——
RWY-ESAnnotations Group=false, Annotations Class=true2023.09—78.9——
ShuffleAcquisition Function=Random, Backbone=BERT2026.05———40.1
ShuffleAcquisition Function=Confidence, Backbone=BERT2026.05———40.3
ShuffleAcquisition Function=CoreSet, Backbone=BERT2026.05———39.1
ShuffleAcquisition Function=BADGE, Backbone=BERT2026.05———40.3