---
pretty_name: Anchor Spatial Reasoning Stage-3 Training Data
language:
- en
license: other
task_categories:
- visual-question-answering
tags:
- spatial-reasoning
- metric-reasoning
- multimodal
configs:
- config_name: master
  data_files:
  - split: train
    path: annotations/*/measurement.parquet
---

# Anchor Spatial Reasoning — Stage 3

This repository is a provenance-preserving collection of spatial **measurement**
questions for answer-supervised training. It is built from VSI-590K,
SpaceVista-Full, HiSpatial-Data, and CA-VQA. SenseNova-SI-8M is intentionally
out of scope for this release.

The first deliverable is the complete `master` collection. Smaller and larger
training views will be derived only after all eligible metric examples have
been retained and audited; they are not early sampling quotas.

## Storage layout

- `annotations/<source>/measurement.parquet`: canonical question/answer rows.
- `manifests/<source>/media_inventory.jsonl`: unique required media and usage.
- `audits/<source>/`: selection counts, validation, checksums, and source policy.
- `media/<source>/*.tar`: uncompressed, approximately 4 GiB media shards.

Media shards retain deterministic member paths. Parquet is the authoritative
annotation format. Source-native answers and provenance are preserved alongside
normalized numeric values and SI values whenever unit conversion is possible.

## Current status

This dataset is being populated source by source. A source is considered
complete only when its annotation row count, media inventory, TAR checksums,
remote file sizes, and missing-media count have all passed validation.

## Licensing and provenance

This is a mixed-source dataset. Every row contains `source`, `source_revision`,
`license`, and `provenance`; downstream users must follow the original source's
terms. In particular, CA-VQA/CubifyAnything carries its own non-commercial
Creative Commons terms. Repository-level `license: other` is intentional and
must not be interpreted as relicensing all source assets under one license.

Source revisions and selection rules are recorded under `audits/` and in every
annotation row. No ownership is asserted over upstream media.