Add Krea2 fixed-seed eval log

This commit is contained in:
2026-06-29 02:49:01 +02:00
parent 40ee843baf
commit 742281f48f
4 changed files with 176 additions and 0 deletions
+40
View File
@@ -61,6 +61,7 @@ import krea_format_route # noqa: E402
import krea_formatter # noqa: E402
import krea_normal_formatter # noqa: E402
import krea_pair_formatter # noqa: E402
import krea2_eval_log # noqa: E402
import krea2_pose_variant_catalog # noqa: E402
import krea_row_fields # noqa: E402
import location_config # noqa: E402
@@ -6803,6 +6804,44 @@ def smoke_krea2_pose_variant_catalog_policy() -> None:
_expect(missing == {}, "Missing pose variant should return an empty mapping")
def smoke_krea2_eval_log_policy() -> None:
log = krea2_eval_log.load_eval_log()
_expect(log.get("version") == 1, "Krea2 eval log version changed unexpectedly")
entries = krea2_eval_log.entries()
_expect(entries, "Krea2 eval log has no entries")
catalog_keys = set(krea2_pose_variant_catalog.variant_keys())
proven_keys = set(krea2_pose_variant_catalog.variant_keys(status="proven"))
accepted_keys = set(krea2_eval_log.variant_keys(result="accepted"))
_expect(proven_keys.issubset(accepted_keys), "Krea2 eval log does not cover every proven pose variant")
seen_ids: set[str] = set()
for entry in entries:
entry_id = _expect_text("krea2_eval_log.entry.id", entry.get("id"), 6)
_expect(entry_id not in seen_ids, f"Krea2 eval log has duplicate entry id {entry_id!r}")
seen_ids.add(entry_id)
variant_key = _expect_text(f"{entry_id}.variant_key", entry.get("variant_key"), 8)
_expect(variant_key in catalog_keys, f"{entry_id} references unknown variant {variant_key!r}")
_expect(isinstance(entry.get("seed"), int), f"{entry_id} has no integer fixed seed")
_expect(entry.get("result") in {"accepted", "rejected", "inconclusive"}, f"{entry_id} has unknown result")
_expect(
entry.get("decision") in {"generator_patch", "prompt_guide_rule", "prompt_only_retry", "needs_more_tests"},
f"{entry_id} has unknown decision",
)
_expect_text(f"{entry_id}.baseline_prompt_summary", entry.get("baseline_prompt_summary"), 20)
_expect_text(f"{entry_id}.candidate_prompt_summary", entry.get("candidate_prompt_summary"), 20)
_expect_text(f"{entry_id}.observation", entry.get("observation"), 30)
for image_key in ("baseline_image", "candidate_image"):
image_path = str(entry.get(image_key) or "")
if image_path:
_expect(Path(image_path).is_absolute(), f"{entry_id}.{image_key} should be absolute when present")
_expect(Path(image_path).is_file(), f"{entry_id}.{image_key} is missing: {image_path}")
boobjob_entries = krea2_eval_log.entries_for_variant("pov_boobjob_upright_cleavage", result="accepted")
_expect(boobjob_entries and boobjob_entries[0].get("seed") == 7302, "Boobjob accepted eval evidence changed")
mutation = krea2_eval_log.entries_for_variant("pov_handjob_upright_centered")[0]
mutation["observation"] = "mutation should not leak"
clean = krea2_eval_log.entries_for_variant("pov_handjob_upright_centered")[0]
_expect(clean.get("observation") != "mutation should not leak", "Krea2 eval log leaked caller mutation")
def smoke_krea_pov_penetration_route() -> None:
pair = pb.build_insta_of_pair(
row_number=1,
@@ -9764,6 +9803,7 @@ SMOKE_CASES: list[tuple[str, Callable[[], None]]] = [
("pov_camera_scene", smoke_pov_camera_scene),
("krea2_pov_pose_variant_catalog", smoke_krea2_pov_pose_variant_catalog),
("krea2_pose_variant_catalog_policy", smoke_krea2_pose_variant_catalog_policy),
("krea2_eval_log_policy", smoke_krea2_eval_log_policy),
("krea_pov_penetration_route", smoke_krea_pov_penetration_route),
("pov_outercourse_position_routes", smoke_pov_outercourse_position_routes),
("pov_oral_position_routes", smoke_pov_oral_position_routes),