README.md

metadata

language:
  - en
tags:
  - code
  - video-evaluation
  - benchmark
  - judge
  - anonymous-release

Anonymous Release — Code

Source code for the benchmark, evaluation, and judge-training pipeline that accompanies the companion data release phyground and the LoRA judge adapter under ../model/.

This drop contains 43 Python source files plus the HTML/CSS/JS assets needed by the human-annotation app. Shell scripts, configuration YAMLs, prompts/answers JSONs, generated dashboards, databases, and binary assets are intentionally excluded — see the dataset and model cards for the artifacts and prompts those scripts consume.

Layout

dataprocessing/
  common/              # Vertex AI / OpenAI client helpers, video-id utilities
  refine/              # Prompt-set construction: enhance, dedup, hard-subset,
                       # humaneval-set assembly, removal sync
  analysis/            # Ablation: prompt-enhancement effect on judges

evals/
  eval_types.py        # Typed result containers for VLM-as-judge runs
  physics_criteria.py  # Physical-law sub-rubric definitions
  sub_questions.py     # Sub-question rendering for CoT/SubQ prompts
  prompts/             # Prompt-template loaders (YAMLs withheld; see model card)
  human_eval/          # Flask-based human-annotation app + tests + templates

judge_training/
  data/                # Build SFT data for ms-swift from raw judgement logs
                       # (schema, sampling, naming, Claude-CoT/DB builders)

Companion artifacts

• Dataset (250 prompts × 8 video models = 2 000 videos, sub-rubric ground truth): ../datasets/ — see its README.md.
• Model (LoRA judge adapter, prompt template, inference script): ../model/ — see its README.md and infer.py.

Dependencies (top-level)

The pipeline relies on the following open-source components. Versions match those reported in the paper.

Component	Used for
transformers, peft, qwen-vl-utils[decord]	Judge inference
ms-swift, deepspeed	Judge LoRA training (ZeRO-2)
vllm (OpenAI-compatible server)	Hosting the base VLM for evaluation
google-genai / Vertex AI	Gemini family runs
anthropic / Vertex AI	Claude family runs
openai Python SDK	OpenAI / GPT family runs
flask, sqlite3, selenium	Human-annotation web app

License

Code is released under the same anonymous-review terms as the rest of this release. No identifying metadata is included.