This is the advanced-only collection. Basic, duplicate source copies, private review archives, generation projects, backups, and caches are not included. One row is one question; cases can contain multiple questions and share media. Source labels and source dataset versions are provenance, not a new blanket review approval.

Loading

There are exactly two configurations, each with 1,199 rows and one unassigned split:

Configuration	Default	File	Contents
`questions`	Yes	`advanced/data/questions.parquet`	Questions, media references, options and response schemas; no answers
`gold`	No	`advanced/data/gold.parquet`	Reference answers, joined to questions by `sample_id`

The JSONL files are equivalent copies, not additional splits. Classification tables, schema maps, quality records, and manifests are metadata, not extra samples. The explicit configuration above prevents these files from being mixed into the two tables.

Install dependencies from the downloaded snapshot:

python -m pip install -r advanced/requirements.txt

Load a local snapshot from its root directory:

from pathlib import Path
from datasets import load_dataset

root = Path(".").resolve()
questions = load_dataset(str(root), split="unassigned")
gold = load_dataset(str(root), "gold", split="unassigned")
answers_by_id = {row["sample_id"]: row["answer_json"] for row in gold}

After publication, replace the placeholder below with the actual dataset repository ID:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo_id = "YOUR_NAMESPACE/YOUR_DATASET"
questions = load_dataset(repo_id, "questions", split="unassigned")
gold = load_dataset(repo_id, "gold", split="unassigned")
row = questions[0]
media_files = [
    hf_hub_download(
        repo_id=repo_id,
        filename="advanced/" + media["path"],
        repo_type="dataset",
    )
    for media in row["media"]
]

Media paths in question rows are relative to advanced/, not the repository root. In a local snapshot, use root / "advanced" / media["path"]. Media paths are ordinary strings, not automatically decoded Hugging Face Video or Image features; table loading does not promise inline video playback in the Dataset Viewer.

Task Categories

Category	Cases	Questions
Forced perspective / real closed-loop judgment	258	258
Spatial alignment	245	245
Containment	97	140
Chirality / rotation equivalence	44	52
Parallax / physical size comparison	21	21
Camera-frame transformation	11	11

The following three categories are reported separately:

Category	Cases	Questions
Assembly, disassembly and next-step prediction	177	215
Structure decomposition and spatial relations	144	156
Maze path understanding and planning	101	101

See classification_summary.json and classification.jsonl. Reporting categories do not replace the original scenario_type or geometry_level. Exact media hashes are checked; semantic deduplication is not claimed. Future splits should respect the retained split_group_id values rather than randomly separating related questions.

Inputs and Answer Contracts

Use question_zh or question_en, all referenced media, options, and response_schema_json as public task inputs. Do not feed gold, source/grouping metadata, or quality records to a model as answer hints. A question may require multiple media files; preserve media_id and option media_ids associations.

Each response is a JSON object with the sole top-level key value. The prompt defines the required types, fields, vocabulary, and array ordering. The independent schema.json response map matches every question-row schema. The packaged scorer implements the declared bilingual aliases and string-normalization extensions; raw JSON Schema validation alone does not implement those extensions.

For the 19-case, 57-question numerical-label-conflict series, numerical dimension annotations are distractors, not true dimensions or calculation inputs. Geometric shapes, relative sizes, and spatial relationships take precedence. Both language prompts state this explicitly. Original answers and motion restrictions are unchanged.

130 questions reference images. Maps, targets, and part references must not be dropped. The 11 camera-pose questions require the last frame before the white interval and the first frame after it; generic sparse sampling or whole-video endpoints are not equivalent inputs.

Scoring and Versions

Prompt protocol: advanced-protocol-20261029-v17.
Active scorer: 2.2-advanced-bilingual.
Quality record: advanced-quality-20261029-v1.
Gold rows retain historical source scorer_version=2.0; this is provenance, not active scorer selection.

From the snapshot root:

python -B verify_release.py
python -B advanced/evaluation/evaluate.py --predictions predictions.jsonl --output scores.json

The output path must not already exist. Each prediction line contains sample_id and prediction_json, which is a JSON string:

{"sample_id":"<sample_id>","prediction_json":"{\"value\":true}"}

This is a serialization example, not a reference answer. Follow the actual question contract. Unknown or duplicate IDs reject the submission. Missing predictions score zero in the full 1,199-question denominator. Primary scores are exact 0/1; auxiliary metrics do not replace the primary score. Record the protocol, scorer version, and manifest hashes with evaluation results.

Review Status and Limits

Case 000007, sample 000007@v001/000007_q01, is passed following the owner's explicit acceptance of its reference answer on 2026-10-29. Its earlier not_passed status is retained as history, not the current decision. Inclusion in the default score is unchanged. See quality_flags.json.

This records an owner answer review, not a new independent visual certification. Technical checks establish file integrity, schema consistency, and deterministic scoring; they do not establish that every question is visually unambiguous, that all answers have independent human agreement, or that synthetic tasks generalize to real-world mechanical physics.

License and Publication Status

Per owner instruction, no reuse license is specified and no LICENSE file is added. No additional rights or reuse permission are asserted here.

This card was prepared in a local release candidate. Its preparation does not create or update a remote repository. Local loading can be tested before upload; remote loading and Dataset Viewer processing must be checked after publication. See README.txt for the candidate checklist and manifest.json for the explicit file scope and checksums.

API Evaluation

A self-contained API evaluation runner is available in advanced/evaluation/. Configure the endpoint/model and environment key, then run:

python3 advanced/evaluation/evaluate.py plan --config advanced/evaluation/api_config.json
python3 advanced/evaluation/evaluate.py run --config advanced/evaluation/api_config.json --output results/model-run

Install the evaluation requirements and ffmpeg/ffprobe first. run calls the configured API and may incur charges. Reports include overall and per-subtask accuracy; gold, media and scorer rules are unchanged.

Downloads last month: 24

Total file size:

2.25 GB