code/object_color/README.md

Object Class / Color prompt-only inference

These source modules and the frozen vocabulary are copied without algorithm changes from winbeau/vbench-repair@7988184. Model weights are paper-final step 300, without development checkpoint selection.

Use the original Qwen environment: Python 3.10+, torch 2.7.1, transformers 4.52.4 and PEFT 0.15.2. Add this directory's src to PYTHONPATH. The base must be a complete local Qwen/Qwen3-8B snapshot at b968826d9c46dd6066d109eabc6255188de91218. Obtain model assets separately and retain their original terms.

from pathlib import Path
from vbench_audit_models.qwen import QwenPromptRouter
from vbench_audit_models.labels import LabelVocabulary

router = QwenPromptRouter.from_local(
    Path("/local/Qwen3-8B"),
    {"object_class": Path("/local/vbench-model/object_class"),
     "color": Path("/local/vbench-model/color")})
vocabulary = LabelVocabulary.from_file(Path("configs/four_dimension/object_color_vocabulary.json"))
compiled = router.compile("color", "a red car", vocabulary, mode="lora")

This compiles only the text conditions. The complete video score still requires the metric's visual backend and frozen protocol. The scripts/ copy is training provenance and needs the full source repository to run.