Background
This is level 2 of 5 and continues the work of the earlier levels. The new files for this level are in workspace/t2/. The records from the start are in workspace/t1/, and the earlier results are in workspace/t1/results/.
Task
Hi,
I've added the published Civil Comments in Context annotations and their documentation. For the paper's reproducibility materials we want an analysis-ready linked release.
Please build it on the corpus work in workspace/t1/results/ and use the primary reporting convention settled there when you reconcile the targets. Where the source alternatives can't be resolved, keep them. Reconstruct the isolated and contextual measurements and the direction of change, and report comment-weighted and rating-weighted results separately.
The release also has to separate observed label switching from statistical evidence that context changed anything. Quantify the conditional no-change distribution of switching, along with the matching expectations for absolute change and squared change. Then assess the evidence for the study as a whole and for individual posts, controlling multiplicity under our standing reporting convention. Make the scope of that sampling benchmark explicit.
The linked annotation CSV and its dataset README go in workspace/t2/results/. Someone else should be able to reproduce the aggregate claims from it and see how sensitive things are at the toxicity boundary, so it needs the per-post evidence and enough source information.
Best, Noor
Input
workspace/t1/: the records supplied with the work (downloads/,research/,tooling/).workspace/t2/: the files that came with level t2 (context_toxicity/).workspace/t1/results/: the results of level t1.
Output
workspace/t2/results/ccc_links.csvworkspace/t2/results/README.md