| ## Background | |
| This is level 1 of 5. The records for this level are in `workspace/t1/`. | |
| ## Task | |
| Hi, | |
| Before we use Civil Comments in the benchmark paper, I need to know what the published splits and targets let us claim about a predictor that sees only the comment text. The full corpus, its card and the reader are in the shared folder. | |
| Please quantify what repeated text and conflicting toxicity targets do to those claims. I'd like the empirical squared-loss floor, and the sharp held-out squared-loss lower bounds and AUROC upper bounds for deterministic real-valued text scores that achieve the smallest possible training squared loss. Show how that training constraint changes the limits, also once training-text overlap is removed. Rows and distinct strings are different populations, so report them separately, and account for the toxicity boundary cases. What I'm after are limits on the recorded benchmark itself. Nobody is proposing a learning method here. | |
| Please also settle the paper's primary reporting agreement from this analysis and record it, because the rest of the paper will be reported on that basis. | |
| The computed results workbook and a reusable corpus-analysis utility with usage documentation go in workspace/t1/results/. Other researchers will extend the utility for the annotation work, so it has to keep recoverable links to the source rows and not just the summary totals. Please follow the lab's research reporting conventions. The discussion that prompted this is in the correspondence folder. | |
| Thanks, | |
| Imani | |
| ## Input | |
| - `workspace/t1/`: the records supplied with the work (`downloads/`, `research/`, `tooling/`). | |
| ## Output | |
| - `workspace/t1/results/corpus_tool` | |
| - `workspace/t1/results/corpus_analysis.xlsx` | |