RAIVNLab / sugar-crepe

[NeurIPS 2023] A faithful benchmark for vision-language compositionality

https://arxiv.org/abs/2306.14610

RAIVNLab/sugar-crepe Issues

Evaluation bug when using GELU vs QuickGELU -- changes the results for the trained models
Updated 5 months ago
Details about human performance
Updated 7 months ago
Access to the hard negatives used during fine-tuning
Closed 8 months ago4
Follow-up question to new NegCLIP variants
Closed 8 months ago5
Table 5 values but broken down into obj, attr, and rel
Closed 8 months ago5
New NegCLIP variants
Closed a year ago3