Skip to content

docs(skill-evals): document grading-schema.json in release-prepare README #964

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
HeaTTap wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −1
Open

docs(skill-evals): document grading-schema.json in release-prepare README #964

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 7 additions & 1 deletion tools/skill-evals/evals/release-prepare/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ uv run --directory tools/skill-evals skill-eval --cli "claude -p" \
evals/release-prepare/step-0-preflight/fixtures/case-1-clean-pass
```

## Grading prose steps (`assertions.json`)
## Grading prose steps (`assertions.json` & `grading-schema.json`)

Steps 1, 2, and 14 emit free-form prose (issue body, PR body), so
their `expected.json` files assert *properties* via `has_*` /
Expand All @@ -44,6 +44,12 @@ deterministic predicate (`regex`, `contains`, `field_true`, `empty`,
`non_empty`) so `--cli` mode grades these cases automatically instead
of reporting `MANUAL`.

Additionally, steps can define a `grading-schema.json` file in their `fixtures/`
directory (see `evals/release-verify-rc/step-3-verify-checksums/fixtures/grading-schema.json` for an example)
which specifies additional `prose_fields` to be evaluated using LLM-based grading.
These custom prose fields extend the default prose field list in `runner.py`, while
all other unspecified fields in `expected.json` default to exact-match evaluation.

Decision fields (title, `proposed`, `category_x_hit`, version
strings) are compared exactly.

Expand Down