# Second annotation round

Purpose: measure how reliable the role labels are by having an independent person
label a sample and comparing.

1. Read `../ANNOTATION_GUIDELINES.md`. Do **not** open `../data/roles/labels.*`,
   `../results/` or any tool output until you have finished.
2. Open `round-2/sample.csv` (150 types, in a fixed random order). For each row,
   follow `url` to the file at its commit and fill `role` with one of
   `entry client service domain persistence event config none`; add a `note` when
   the decision is not obvious.
3. Save the file and run `python3 scripts/kappa.py annotation/round-2/sample.csv`
   from the repository root. It prints the agreement, Cohen's kappa and the
   disagreements.
4. The two annotators discuss the disagreements; corrected labels and the
   agreement figures go into the next release (see CHANGELOG.md).

The sample is drawn in proportion to each project's labels, by a seeded hash, so it
can be regenerated with `node build/build.js roles`.