In toxicity detection, given a text, the goal is to predict whether the text has toxic content.

toxicity_detection

How many models this model outperform on average (over columns).

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Quasi-exact match: Fraction of instances that the predicted output matches a correct reference up to light processing.

accuracy

min=0.03, mean=0.553, max=0.968, sum=29.863 (54)

min=0, mean=0.532, max=0.996, sum=28.713 (54)

min=0.014, mean=0.529, max=0.991, sum=28.55 (54)

min=0.008, mean=0.546, max=1, sum=29.501 (54)

min=0.011, mean=0.57, max=1, sum=30.805 (54)

min=0.011, mean=0.547, max=0.998, sum=29.525 (54)

min=0.149, mean=0.57, max=0.909, sum=30.8 (54)

min=0.003, mean=0.544, max=1, sum=29.372 (54)

min=0.014, mean=0.524, max=0.997, sum=28.276 (54)

min=0.049, mean=0.562, max=0.984, sum=30.331 (54)

min=0.182, mean=0.61, max=0.939, sum=32.915 (54)

min=0.293, mean=0.62, max=0.92, sum=33.467 (54)

min=0, mean=0.234, max=0.985, sum=12.634 (54)

min=0.001, mean=0.532, max=1, sum=28.726 (54)

min=0, mean=0.507, max=1, sum=27.395 (54)

min=0, mean=0.504, max=1, sum=27.205 (54)

min=0, mean=0.501, max=1, sum=27.062 (54)

min=0.035, mean=0.524, max=0.968, sum=28.319 (54)

min=0.009, mean=0.54, max=1, sum=29.17 (54)

min=0.254, mean=0.601, max=0.86, sum=32.478 (54)

min=0.002, mean=0.52, max=1, sum=28.06 (54)

min=0, mean=0.516, max=1, sum=27.878 (54)

min=0, mean=0.509, max=0.998, sum=27.462 (54)

min=0, mean=0.521, max=1, sum=28.146 (54)

min=0, mean=0.505, max=1, sum=27.251 (54)

min=0, mean=0.506, max=1, sum=27.302 (54)

min=0.171, mean=0.601, max=0.983, sum=32.472 (54)

min=0.053, mean=0.532, max=0.955, sum=28.701 (54)

min=0.006, mean=0.532, max=1, sum=28.723 (54)

min=0.012, mean=0.539, max=1, sum=29.083 (54)

min=0.005, mean=0.519, max=0.996, sum=28.025 (54)

min=0.52, mean=0.684, max=0.863, sum=36.959 (54)

min=0.4, mean=0.668, max=0.876, sum=36.093 (54)

min=0.04, mean=0.537, max=0.93, sum=29.013 (54)

min=0, mean=0.499, max=1, sum=26.951 (54)

min=0.434, mean=0.674, max=0.824, sum=36.394 (54)

min=0.013, mean=0.549, max=0.996, sum=9.877 (18)

min=0, mean=0.555, max=0.877, sum=29.98 (54)

min=0, mean=0.008, max=0.344, sum=0.406 (54)

Accuracy

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

10-bin expected calibration error: The average difference between the model's confidence and accuracy, averaged across 10 bins where each bin contains an equal number of points (only computed for classification tasks). Warning - not reliable for small datasets (e.g., with < 300 examples) because each bin will have very few examples.

calibration

min=0.048, mean=0.27, max=0.587, sum=14.569 (54)

min=0.073, mean=0.377, max=0.573, sum=20.347 (54)

min=0.228, mean=0.408, max=0.593, sum=22.008 (54)

min=0.131, mean=0.376, max=0.649, sum=20.307 (54)

min=0.07, mean=0.314, max=0.578, sum=16.963 (54)

min=0.092, mean=0.381, max=0.605, sum=20.555 (54)

min=0.037, mean=0.19, max=0.41, sum=10.254 (54)

min=0.071, mean=0.28, max=0.632, sum=15.102 (54)

min=0.112, mean=0.359, max=0.619, sum=19.409 (54)

min=0.051, mean=0.272, max=0.563, sum=14.71 (54)

min=0.048, mean=0.179, max=0.449, sum=9.655 (54)

min=0.069, mean=0.262, max=0.456, sum=14.142 (54)

min=0.067, mean=0.308, max=0.574, sum=16.631 (54)

min=0.051, mean=0.327, max=0.708, sum=17.639 (54)

min=0.1, mean=0.384, max=0.705, sum=20.717 (54)

min=0.176, mean=0.459, max=0.641, sum=24.77 (54)

min=0.201, mean=0.486, max=0.8, sum=26.269 (54)

min=0.056, mean=0.313, max=0.651, sum=16.899 (54)

min=0.265, mean=0.487, max=0.736, sum=26.317 (54)

min=0.113, mean=0.358, max=0.735, sum=19.322 (54)

min=0.054, mean=0.161, max=0.416, sum=8.676 (54)

min=0.075, mean=0.409, max=0.626, sum=22.076 (54)

min=0.26, mean=0.444, max=0.593, sum=23.994 (54)

min=0.108, mean=0.38, max=0.553, sum=20.519 (54)

min=0.123, mean=0.404, max=0.585, sum=21.802 (54)

min=0.226, mean=0.462, max=0.633, sum=24.957 (54)

min=0.299, mean=0.474, max=0.666, sum=25.591 (54)

min=0.058, mean=0.213, max=0.447, sum=11.516 (54)

min=0.052, mean=0.248, max=0.54, sum=13.38 (54)

min=0.083, mean=0.396, max=0.664, sum=21.389 (54)

min=0.042, mean=0.293, max=0.601, sum=15.826 (54)

min=0.063, mean=0.31, max=0.598, sum=16.723 (54)

min=0.096, mean=0.355, max=0.704, sum=19.19 (54)

min=0.12, mean=0.292, max=0.449, sum=15.772 (54)

min=0.018, mean=0.183, max=0.424, sum=9.875 (54)

min=0.049, mean=0.262, max=0.674, sum=14.15 (54)

min=0.092, mean=0.499, max=0.911, sum=26.966 (54)

min=0.012, mean=0.479, max=0.985, sum=25.845 (54)

min=0.176, mean=0.326, max=0.566, sum=17.606 (54)

min=0.105, mean=0.303, max=0.532, sum=5.455 (18)

min=0.22, mean=0.486, max=0.749, sum=26.268 (54)

min=0.123, mean=0.445, max=1, sum=24.02 (54)

min=0.656, mean=0.992, max=1, sum=53.594 (54)

min=0.108, mean=0.437, max=0.784, sum=23.581 (54)

Calibration

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Quasi-exact match: Fraction of instances that the predicted output matches a correct reference up to light processing.
- Perturbation Robustness: Computes worst case over different robustness perturbations (misspellings, formatting, contrast sets).

robustness

min=0.027, mean=0.271, max=0.732, sum=14.649 (54)

min=0, mean=0.444, max=0.984, sum=23.966 (54)

min=0.014, mean=0.417, max=0.938, sum=22.51 (54)

min=0, mean=0.495, max=0.995, sum=26.738 (54)

min=0.009, mean=0.449, max=0.979, sum=24.224 (54)

min=0, mean=0.488, max=0.986, sum=26.326 (54)

min=0.116, mean=0.469, max=0.844, sum=25.305 (54)

min=0, mean=0.416, max=0.99, sum=22.479 (54)

min=0.011, mean=0.368, max=0.874, sum=19.881 (54)

min=0.035, mean=0.263, max=0.67, sum=14.178 (54)

min=0.105, mean=0.514, max=0.854, sum=27.755 (54)

min=0.088, mean=0.467, max=0.827, sum=25.192 (54)

min=0, mean=0.087, max=0.824, sum=4.704 (54)

min=0, mean=0.32, max=0.817, sum=17.265 (54)

min=0, mean=0.333, max=0.95, sum=17.981 (54)

min=0, mean=0.136, max=0.736, sum=7.362 (54)

min=0, mean=0.434, max=1, sum=23.451 (54)

min=0.012, mean=0.408, max=0.908, sum=22.047 (54)

min=0, mean=0.353, max=0.931, sum=19.089 (54)

min=0.205, mean=0.535, max=0.84, sum=28.866 (54)

min=0, mean=0.418, max=1, sum=22.597 (54)

min=0, mean=0.392, max=0.991, sum=21.175 (54)

min=0, mean=0.45, max=0.983, sum=24.293 (54)

min=0, mean=0.184, max=0.769, sum=9.952 (54)

min=0, mean=0.305, max=0.939, sum=16.459 (54)

min=0.069, mean=0.409, max=0.689, sum=22.106 (54)

min=0.022, mean=0.336, max=0.831, sum=18.169 (54)

min=0, mean=0.461, max=1, sum=24.899 (54)

min=0.002, mean=0.347, max=1, sum=18.748 (54)

min=0, mean=0.4, max=0.996, sum=21.618 (54)

min=0, mean=0.421, max=1, sum=22.752 (54)

min=0.366, mean=0.594, max=0.838, sum=32.08 (54)

min=0.287, mean=0.567, max=0.838, sum=30.64 (54)

min=0, mean=0.129, max=0.39, sum=6.954 (54)

min=0, mean=0.491, max=1, sum=26.518 (54)

min=0.313, mean=0.604, max=0.824, sum=32.639 (54)

min=0, mean=0.336, max=0.996, sum=6.045 (18)

min=0, mean=0.444, max=0.774, sum=23.954 (54)

min=0, mean=0.006, max=0.319, sum=0.347 (54)

min=0, mean=0.463, max=1, sum=25.008 (54)

Robustness

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Quasi-exact match: Fraction of instances that the predicted output matches a correct reference up to light processing.
- Perturbation Fairness: Computes worst case over different fairness perturbations (changing dialect, race of names, gender).

fairness

min=0.006, mean=0.478, max=0.958, sum=25.823 (54)

min=0, mean=0.447, max=0.962, sum=24.127 (54)

min=0.014, mean=0.482, max=0.962, sum=26.023 (54)

min=0.005, mean=0.404, max=0.901, sum=21.814 (54)

min=0.005, mean=0.507, max=0.995, sum=27.37 (54)

min=0, mean=0.445, max=0.973, sum=24.007 (54)

min=0.143, mean=0.403, max=0.834, sum=21.752 (54)

min=0, mean=0.397, max=1, sum=21.425 (54)

min=0, mean=0.462, max=0.985, sum=24.963 (54)

min=0.014, mean=0.432, max=0.912, sum=23.313 (54)

min=0.105, mean=0.512, max=0.939, sum=27.636 (54)

min=0.252, mean=0.546, max=0.91, sum=29.488 (54)

min=0, mean=0.165, max=0.947, sum=8.894 (54)

min=0.001, mean=0.479, max=1, sum=25.855 (54)

min=0, mean=0.443, max=1, sum=23.917 (54)

min=0, mean=0.489, max=1, sum=26.387 (54)

min=0, mean=0.495, max=1, sum=26.744 (54)

min=0.03, mean=0.415, max=0.875, sum=22.43 (54)

min=0, mean=0.493, max=1, sum=26.609 (54)

min=0.002, mean=0.496, max=1, sum=26.757 (54)

min=0.222, mean=0.544, max=0.85, sum=29.397 (54)

min=0, mean=0.488, max=1, sum=26.356 (54)

min=0, mean=0.491, max=1, sum=26.497 (54)

min=0, mean=0.329, max=0.991, sum=17.759 (54)

min=0, mean=0.423, max=0.975, sum=22.816 (54)

min=0, mean=0.491, max=1, sum=26.489 (54)

min=0.047, mean=0.48, max=0.97, sum=25.944 (54)

min=0.042, mean=0.473, max=0.947, sum=25.533 (54)

min=0.003, mean=0.478, max=1, sum=25.83 (54)

min=0, mean=0.412, max=1, sum=22.222 (54)

min=0, mean=0.474, max=0.994, sum=25.57 (54)

min=0, mean=0.436, max=1, sum=23.537 (54)

min=0.253, mean=0.559, max=0.863, sum=30.179 (54)

min=0.082, mean=0.463, max=0.851, sum=24.991 (54)

min=0.02, mean=0.471, max=0.874, sum=25.434 (54)

min=0, mean=0.499, max=1, sum=26.936 (54)

min=0.024, mean=0.433, max=0.824, sum=23.4 (54)

min=0, mean=0.393, max=0.996, sum=7.082 (18)

min=0, mean=0.449, max=0.871, sum=24.224 (54)

min=0, mean=0.006, max=0.288, sum=0.338 (54)

min=0, mean=0.456, max=0.998, sum=24.603 (54)

Fairness

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Stereotypical associations (race, profession): Measures uneven association of racial groups (Asian, Hispanic, White) with target professions. This measurement is based on cooccurence statistics between the racially-associated names (word list from [Garg et al., 2018](https://www.pnas.org/doi/10.1073/pnas.1720347115); race associations based on US Census statistics) and the target professions (word list from [Bolukbasi et al., 2016](https://papers.nips.cc/paper/2016/hash/a486cd07e4ac3d270571622f4f316ec5-Abstract.html)).

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Stereotypical associations (gender, profession): Measures uneven association of gender groups (male, female) with target professions. This measurement is based on cooccurence statistics between the gender terms (word list from [Bolukbasi et al., 2016](https://papers.nips.cc/paper/2016/hash/a486cd07e4ac3d270571622f4f316ec5-Abstract.html)) and the target professions (word list from [Bolukbasi et al., 2016](https://papers.nips.cc/paper/2016/hash/a486cd07e4ac3d270571622f4f316ec5-Abstract.html)).

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Demographic representation (race): Measures uneven representation of racial groups (Asian, Hispanic, White). This measurement is based on disparities in the frequency statistics across racially-associated names (word list from [Garg et al., 2018](https://www.pnas.org/doi/10.1073/pnas.1720347115); race associations based on US Census statistics).

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Demographic representation (gender): Measures uneven representation of gender groups (male, female). This measurement is based on disparities in the frequency statistics across gender terms (word list from [Bolukbasi et al., 2016](https://papers.nips.cc/paper/2016/hash/a486cd07e4ac3d270571622f4f316ec5-Abstract.html)).

bias

min=0.3, mean=0.459, max=0.5, sum=5.503 (12)

min=0.333, mean=0.5, max=0.667, sum=1 (2)

Bias

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Toxic fraction: Fraction of model outputs that are toxic (based on the PerspectiveAPI toxicity classifier).

toxicity

min=0, mean=0.0, max=0.008, sum=0.025 (54)

min=0, mean=0.0, max=0.001, sum=0.001 (54)

Toxicity

The CivilComments benchmark for toxicity detection [(Borkan et al., 2019)](https://arxiv.org/pdf/1903.04561.pdf).

Denoised inference runtime (s): Average time to process a request to the model minus performance contention by using profiled runtimes from multiple trials of SyntheticEfficiencyScenario.

efficiency

min=0.43, mean=0.552, max=0.724, sum=29.829 (54)

min=0.338, mean=0.434, max=0.564, sum=23.454 (54)

min=0.418, mean=0.482, max=0.621, sum=26.002 (54)

min=0.555, mean=0.594, max=0.756, sum=32.071 (54)

min=0.316, mean=0.533, max=1.372, sum=28.76 (54)

min=0.328, mean=0.391, max=0.487, sum=21.126 (54)

min=0.464, mean=0.546, max=0.711, sum=29.484 (54)