Background
This is level 5 of 5 and continues the work of the earlier levels. The new files for this level are in workspace/t5/. The records from the start are in workspace/t1/, and the earlier results are in workspace/t1/results/ to workspace/t4/results/.
Task
Hi,
We've picked the toxic-bert Hub artifact as the model for the benchmark reproduction. Its weights, tokenizer assets, configuration and card are now in the folder.
Please run it on the cross-campaign cohorts in workspace/t4/results/, including the linkage sensitivities worked out there. Drawing on the earlier annotation analyses, quantify how its apparent error changes with context and with rater population. We need empirical squared-error and AUROC results, the finite-panel-adjusted comparison, and sharp attainable ranges over completions of the unresolved ratings. Please carry the measurement choices from the earlier analyses and simultaneous uncertainty through all of these results, including across the linkage alternatives.
The per-comment predictions and the validation report go in workspace/t5/results/. The report should include the cohort accounting that ties these results to the supplementary workbook, say which artifact was actually reproduced, and state which model-performance claims stay stable under these choices. It goes straight into the research paper.
Thanks, Imani
Input
workspace/t1/: the records supplied with the work (downloads/,research/,tooling/).workspace/t2/: the files that came with level t2 (context_toxicity/).workspace/t3/: the files that came with level t3 (context_toxicity/,wiki-detox/).workspace/t4/: the files that came with level t4 (jigsaw-specialized-rater-pools-dataset/,research/).workspace/t5/: the files that came with level t5 (models/).workspace/t1/results/: the results of level t1.workspace/t2/results/: the results of level t2.workspace/t3/results/: the results of level t3.workspace/t4/results/: the results of level t4.
Output
workspace/t5/results/bridge_predictions.csvworkspace/t5/results/artifact_validation.pdf