VMem-Bench / trackB /scripts /screenplay_to_gt_source.py
Suchenl's picture
Initial VMem-Bench gold JSON and prompts (no source videos) (part 2)
55c86fd verified
Raw
History Blame Contribute Delete
7.36 kB
#!/usr/bin/env python3
"""Convert MemStrata production screenplays into Track B v2 gt_source files.
This is a one-way bootstrap helper for upgrading the old Track B samples. It
keeps the benchmark-facing author layer small: entities, state machines,
lookalike pairs, scenes, per-shot present sets, and semantic events.
"""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
from typing import Any
HERE = Path(__file__).resolve().parent.parent # trackB root (scripts/ lives one level down)
REPO = HERE.parents[3]
SCREENPLAY_DIR = REPO / "methods" / "MemStrata" / "production" / "screenplay" / "products" / "cn"
OUT_DIR = HERE / "gt_source"
EID_TAG_RE = re.compile(r"[((]E\d+[))]")
def _kind(entity_type: str) -> str:
if entity_type == "object":
return "prop"
return entity_type
def _clean_action(text: str) -> str:
return EID_TAG_RE.sub("", text).replace(" ", " ").strip()
def _normalize_entity_names(story_id: str, text: str) -> str:
aliases = {
"0002_night_market_courier": {
"木盒": "封蜡木盒",
},
"0003_desert_archaeologist": {
"泥碑": "楔形泥碑",
"羊皮地图": "伪造羊皮地图",
},
}
for src, dst in aliases.get(story_id, {}).items():
if dst not in text:
text = text.replace(src, dst)
return text
def _state_label(raw: str) -> str:
labels = {
"pried": "被撬开",
"empty": "空盒",
"shattered": "碎裂",
"reassembled": "拼合复原",
"discredited": "证伪",
"burned": "焚毁",
}
return labels.get(raw, raw)
def _initial_label(story_id: str, eid: str) -> str:
labels = {
("0002_night_market_courier", "E3"): "完好封蜡",
("0003_desert_archaeologist", "E2"): "完好",
("0003_desert_archaeologist", "E3"): "伪图",
}
return labels.get((story_id, eid), "初始")
def _state_changes(overrides: dict[str, Any]) -> dict[str, dict[str, str]]:
by_shot: dict[str, dict[str, str]] = {}
for item in overrides.get("state_changes", []):
by_shot.setdefault(item["from_shot"], {})[item["eid"]] = _state_label(item["label"])
return by_shot
def _state_defs(story_id: str, screenplay: dict[str, Any], overrides: dict[str, Any]) -> tuple[dict[str, Any], dict[str, list[str]]]:
changed: dict[str, list[tuple[str, str]]] = {}
for item in overrides.get("state_changes", []):
changed.setdefault(item["eid"], []).append((_state_label(item["label"]), item["desc"]))
entities: dict[str, Any] = {}
machines: dict[str, list[str]] = {}
for ent in screenplay["main_entities"]:
eid = ent["entity_id"]
base = {"name": ent["name"], "kind": _kind(ent["entity_type"])}
if eid in changed:
initial = _initial_label(story_id, eid)
states = {initial: ent["appearance"]}
for label, desc in changed[eid]:
states[label] = desc
base["states"] = states
machines[eid] = list(states)
else:
base["appearance"] = ent["appearance"]
entities[eid] = base
return entities, machines
def _shot_present(shot: dict[str, Any]) -> list[str]:
present: list[str] = []
for aid in shot.get("active_characters", []):
if aid not in present:
present.append(aid)
for asset in shot.get("planned_assets", []):
eid = asset.get("planned_asset_id")
if not eid or asset.get("operation") == "avoid":
continue
if asset.get("required", True) and eid not in present:
present.append(eid)
return present
def _lookalike_present(pairs: list[dict[str, Any]], present: list[str]) -> dict[str, Any] | None:
present_set = set(present)
for pair in pairs:
members = [eid for eid in pair["pair"] if eid in present_set]
if members:
return {"pair": pair["pair"], "members": members}
return None
def convert(story_id: str) -> dict[str, Any]:
screenplay_path = SCREENPLAY_DIR / f"{story_id}.json"
overrides_path = SCREENPLAY_DIR / f"{story_id}.overrides.json"
screenplay = json.loads(screenplay_path.read_text(encoding="utf-8"))
overrides = json.loads(overrides_path.read_text(encoding="utf-8"))
entities, state_machines = _state_defs(story_id, screenplay, overrides)
shot_changes = _state_changes(overrides)
lookalike_pairs = overrides.get("lookalike_pairs", [])
shots_by_scene: dict[str, list[dict[str, Any]]] = {}
for shot in screenplay["production_screenplay"]["shots"]:
shots_by_scene.setdefault(shot["scene_id"], []).append(shot)
scenes: list[dict[str, Any]] = []
for scene in screenplay["production_screenplay"]["scenes"]:
scene_shots = shots_by_scene.get(scene["scene_id"], [])
actions: list[Any] = []
for shot in scene_shots:
action = _normalize_entity_names(
story_id,
_clean_action(" ".join(shot.get("visual_track", {}).get("actions", []))),
)
present = _shot_present(shot)
item: dict[str, Any] = {"action": action, "present": present}
events: list[dict[str, Any]] = []
for eid, label in shot_changes.get(shot["shot_id"], {}).items():
event = {"type": "state_change", "eid": eid, "to": label}
if label == "焚毁":
event = {"type": "remove", "eid": eid, "to": label, "reason": "destroyed", "shown": True}
events.append(event)
if events:
item["events"] = events
lp = _lookalike_present(lookalike_pairs, present)
if lp:
item["lookalike_present"] = lp
actions.append(item)
scenes.append({
"id": scene["scene_id"],
"setting": scene.get("scene_title", scene["scene_id"]),
"present": scene.get("entities_present", []),
"actions": actions,
})
return {
"story_id": story_id,
"title": screenplay.get("story_name", story_id),
"premise": screenplay.get("story_overview", ""),
"_comment": "Generated from MemStrata production screenplay plus overrides; hand-audit before freezing.",
"segment_sec": 5.0,
"gap_long_threshold": 30,
"avoidance_probe_window": 4,
"probe_target_default": 5,
"entities": entities,
"state_machines": state_machines,
"lookalike_pairs": lookalike_pairs,
"scenes": scenes,
}
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("stories", nargs="+")
parser.add_argument("--out-dir", type=Path, default=OUT_DIR)
parser.add_argument("--overwrite", action="store_true")
args = parser.parse_args()
args.out_dir.mkdir(parents=True, exist_ok=True)
for story_id in args.stories:
out = args.out_dir / f"{story_id}.json"
if out.exists() and not args.overwrite:
raise SystemExit(f"{out} exists; pass --overwrite to replace it")
data = convert(story_id)
out.write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
print(f"wrote {out}")
if __name__ == "__main__":
main()