README.md

---
license: other
license_name: mixed-per-image-licences
task_categories:
- image-text-to-text
language:
- th
size_categories:
- 10K<n<100K
pretty_name: Thai Spatial Reasoning
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
  - split: test
    path: data/test-*.parquet
---

# Thai Spatial Reasoning 1.0.0

Thai spatial captions and question-answer pairs for continual pretraining of a Thai
foundation VLM. Synthetic means the Thai text and the spatial annotations are generated or
processed; the primary images are real photographs.

Images are **not** covered by one blanket licence, so the card does not point at one file:
every image has its own licence, creator, and attribution in
[`rights/attribution.csv`](rights/attribution.csv) and
[`rights/ledger.parquet`](rights/ledger.parquet). The dataset-level `other` licence applies to
the annotations, the generated text, and the rendered images. (The Hub rejects a relative
`license_link`; the ledger inside the repository is the licence statement.)

## Contents

| Item | Value |
| --- | --- |
| Images | 34764 |
| Logical examples (CPT) | 243348 |
| Per image | 1 caption + 4 question-answer pairs |
| Splits | {"train": 30481, "validation": 2271, "test": 2012} |
| Sources | real photographs with rights evidence, plus controlled counterfactual renders |

`data/*.parquet` is the canonical table: one row per image, with an `Image` column, one
caption, a list of four question-answer structures, the relation list, and the rights
fields. `cpt/examples-*.parquet` is the text-only index that expands one image row into
five logical examples, so image bytes are stored once.

## Method

1. Real images come from a source with object boxes. Rights are checked per image against a
   licence allowlist; a verification date and an evidence URL are recorded for each one.
   Images with person labels wait for a human privacy decision and do not ship without it.
2. Splits are assigned per duplicate group and per rendered scene family before any text
   exists. No group crosses a split boundary.
3. Facts come from conservative image-frame geometry over boxes (a real separation on one
   axis and real alignment on the other, no clipped axis, no occluded reference object) and
   from source annotations for containment. Contact, depth, and distance are never inferred
   from 2-D boxes; for the renders they are measured from known 3-D state.
4. Thai text is assembled from those facts with deterministic templates, so each sentence
   traces to fact ids, and the factual meaning does not depend on a model.
5. An object whose class repeats in an image is named by what it looks like before any position is
   used ("คนที่ใส่เสื้อสีเขียว", not "คนที่อยู่ทางซ้าย"). A pinned vision model returns one value per
   field — colour, garment, pose, material — from a closed list; it writes no Thai, and an
   attribute can neither create nor remove a spatial fact. Renders are never sent to it, because
   their colours are known from the scene state.
6. Every relation and every sentence is checked against its source fact, and a
   Thai-speaking review sample of 0 examples is scored before
   release.

## Evaluation and quality

**The Thai-speaking review gate is still open for this build.** A sample of examples are queued in `reports/audit_queue.csv` for a native-speaker review of the relation, the language, and the faithfulness. This build ships before those verdicts exist, and it will be re-packaged with them; treat the relation labels as machine-verified only.

Validation report: [`reports/validation.json`](reports/validation.json). Audit report:
[`reports/audit.json`](reports/audit.json). Splits:
[`reports/split_manifest.json`](reports/split_manifest.json). Release gates checked:
rights_complete=pass, privacy_review_complete=pass, split_integrity=pass, logical_examples_per_image=pass, artifact_integrity=pass, human_audit=fail, lexicon_review=fail, declared_scale=fail, release_size=pass.

Real-image acceptance in this build: {'real': {'accepted': 5778, 'images': 20823, 'rate': 0.2775}, 'rendered': {'accepted': 28986, 'images': 40000, 'rate': 0.7247}}.
Relations available: {'above': 33571, 'behind': 177397, 'below': 33571, 'far': 4975, 'in_front_of': 177397, 'larger_than': 104893, 'left_of': 46479, 'near': 81090, 'not_touching': 199465, 'occludes': 25414, 'right_of': 46479, 'smaller_than': 104893, 'touching': 9364}.

## Intended use

Continual pretraining and evaluation of Thai vision-language models on spatial language:
left/right, above/below, in front of/behind, near/far, contact, size comparison, and
containment, in image, object, and world reference frames.

## Limitations

* Image-frame left/right and above/below are reliable only where the boxes are clearly
  separated and aligned; the dataset does not claim 3-D layout from photographs.
* Renderer output is flat-shaded and synthetic: it is a controlled probe for relation
  contrasts, not a substitute for photographs.
* Relation balance reflects what the source annotations support. Under-represented
  relations are supplemented by renders, not invented for photographs.
* Thai text is template generated.
* Descriptions such as `เสื้อสีเขียว` come from a pinned vision model, not from the source
  annotation. A spot check of six objects found five correct; an attribute is only used when it
  tells one object apart from the others in the same image, so an ambiguous description is dropped
  rather than applied to the wrong object. Treat the attribute text as model-generated.

## Version and citation

Version 1.0.0, built by run `release-v13` with config `thai_spatial_reasoning_v1.yaml`
(config hash `dff750b131deec59`, code `ad8324a6dcd67a2a0780b05eb0f139108f3a19a3`). See the project's `CITATIONS.md` for
the methodology references.

## Removal and contact

To request removal of an image or to report a rights problem, open an issue in the
project repository or contact the dataset maintainers (smartwhatt/thai_spatial_reasoning). Each
image row carries `source_url`, `image_license`, and `rights_evidence_uri`, which is enough
to identify and remove the exact image in a future revision.

## Token budget

Measured with the `Qwen/Qwen3-VL-2B-Instruct` image processor over every one of the 34,764 images in this release,
so these are the tokens a Qwen3-VL model actually sees, not an estimate. Text is counted with the
processor tokenizer over captions, questions and answers.

| split | images | visual tokens | text tokens |
|---|---|---|---|
| train | 30,481 | 13,715,675 | 6,628,004 |
| validation | 2,271 | 3,788,003 | 683,998 |
| test | 2,012 | 1,467,105 | 481,804 |
| **total** | **34,764** | **18,970,783** | **7,793,806** |

**Total: 26,764,589 tokens** — 18,970,783 visual plus 7,793,806 text, an average of 545 visual and
224 text tokens per image.

Real images cost far more than renders: 1,778 visual tokens per image on average against
300 for a render, because a 1240x1754 photograph is not a 640x480 scene. The training
splits are render-heavy, so most of the training budget today is the fixed 300-token renders, and the
real photographs are where the per-image weight sits.

The continuous-pretraining index is a separate figure: 243,348 rows carrying 8,175,985 text
tokens, and about **140,800,645 tokens** if every row trains with its image attached. Quote
that number when a training recipe asks for a slice size.