Training Time for Concept Learning on AG News
234Training TimeClassifSAE (K=10)
Evaluation Results
| Method | Links | |
|---|---|---|
| ClassifSAE (K=10)Classifier Model=BERT(110M)2025.06 | 234 | |
| ClassifSAE (K=10)Classifier Model=Pythia(410M)2025.06 | 253 | |
| ClassifSAE (K=10)Classifier Model=DeBERTa-v3(304M)2025.06 | 254 | |
| ClassifSAE (K=10)Classifier Model=Pythia(1B)2025.06 | 501 | |
| ConceptSHAPClassifier Model=BERT(110M)2025.06 | 910 | |
| ConceptSHAPClassifier Model=DeBERTa-v3(304M)2025.06 | 950 | |
| HI-ConceptClassifier Model=BERT(110M)2025.06 | 1,056 | |
| HI-ConceptClassifier Model=DeBERTa-v3(304M)2025.06 | 1,150 | |
| ClassifSAE (K=10)Classifier Model=GPT-J(6B)2025.06 | 1,347 | |
| ClassifSAE (K=10)Classifier Model=Mistral-Inst.(7B)2025.06 | 1,401 | |
| ClassifSAE (K=10)Classifier Model=Llama-Inst.(8B)2025.06 | 1,408 | |
| ConceptSHAPClassifier Model=Pythia(410M)2025.06 | 1,914 | |
| ConceptSHAPClassifier Model=Pythia(1B)2025.06 | 2,339 | |
| HI-ConceptClassifier Model=Pythia(410M)2025.06 | 2,610 | |
| ConceptSHAPClassifier Model=Mistral-Inst.(7B)2025.06 | 2,853 | |
| HI-ConceptClassifier Model=Pythia(1B)2025.06 | 3,030 | |
| ConceptSHAPClassifier Model=GPT-J(6B)2025.06 | 3,247 | |
| ConceptSHAPClassifier Model=Llama-Inst.(8B)2025.06 | 4,538 | |
| HI-ConceptClassifier Model=Mistral-Inst.(7B)2025.06 | 4,952 | |
| HI-ConceptClassifier Model=GPT-J(6B)2025.06 | 5,541 | |
| HI-ConceptClassifier Model=Llama-Inst.(8B)2025.06 | 12,347 |