Share your thoughts, 1 month free Claude Pro on usSee more

Performance prediction on MNLI source domains (in-domain)

0.699ROC AUC

Cosine distance (fine-tuned)

Updated 1mo ago

Evaluation Results

Method	Links
Cosine distance (fine-tuned) 2023.05		0.699
Vocabulary, structural, semantic drift 2023.05		0.525
Combined prev. model-agnostic 2023.05		0.52
Structural drift 2023.05		0.516
Semantic drift 2023.05		0.516
Baseline (no-performance drop) 2023.05		0.5
Token frequency cross-entropy 2023.05		0.5
Token frequency JS-div 2023.05		0.496
Cosine distance (pre-trained) 2023.05		0.484
Vocabulary drift 2023.05		0.474