3D spatial reasoning: human-approved cases

This snapshot contains 2,616 current cases and 2,976 questions, with 2,802 videos and 99 images. It includes the current versions marked as passed in the local human review system. Failed and uncertain cases are excluded. Historical approved versions are kept separately by the data owner and are not part of this snapshot.

Standard table

data/questions.parquet is the default dataset entry point. Each row is one question and preserves all 18 standard fields. data/questions.jsonl has identical decoded records. The data schema version is 1.0 and the scorer version is 2.0. All rows have split=unassigned; no training/test split has been assigned.

from datasets import load_dataset
questions = load_dataset("flymanss/3D", split="unassigned")

For a private repository, authenticate with Hugging Face before loading. To load a downloaded checkout directly:

questions = load_dataset(
    "parquet", data_files={"unassigned": "data/questions.parquet"}, split="unassigned"
)

Media paths are relative to the repository root. Files in media/ are the original video and image bytes, named by SHA-256. The standard table stores media records and paths, not Hugging Face Video features. Multi-video questions, reference images, and option-to-media relationships are preserved.

English preview and responses

After downloading the repository, open viewer/index.html in a browser to view English questions with their media. The preview does not display category headings or answers. viewer/english.jsonl is a derived view retaining identifiers, classification, scoring fields, and complete media/options references. The standard mother table retains the source Chinese and English questions.

Responses must use the strict {"value": ...} JSON wrapper defined in each row's response_schema_json. Original answer-option IDs are preserved, including historical non-English IDs. evaluation/scorer.py implements exact scoring, but private gold answers are required to run an actual evaluation. Gold answers, human-review notes, and private source evidence are not included in this repository.

Validation and known source issues

The local snapshot passed standard-schema checks, JSONL/Parquet equivalence, local Hugging Face loading, source-answer replay, invalid-format tests, media hash checks, and full video/image decoding. Answer replay was a conversion self-test, not a model benchmark result. Missing geometry levels for 667 cases were filled using an explicitly confirmed mapping; existing source levels were preserved. Subtasks remain null where no source subdivision was recorded.

Four human-approved questions name all components, including D/E, but later refer to an A/B/C answer order. Their original wording and review decisions are preserved; their response mapping includes all publicly named components:

• local_fancy_700_20261019_06_0155@attempt_01/q1
• local_fancy_700_20261019_06_0156@attempt_01/q1
• local_fancy_700_20261019_06_0223@attempt_01/q1
• local_fancy_700_20261019_06_0346@attempt_01/q1

No license is specified by this dataset card. The uploader has not added a new license or changed the source review decisions.