AbstractPhil/thought-symbolic-image-captions

Thought-Symbolic Image Captions

A text-only dataset of thinking about images that nobody sees. Every row holds a hidden image specification (a small set of records in a fixed vocabulary), a chain of thoughts about that image at a stated depth, the wordings an artist would use for it (a caption, a generation prompt, tags, a critique), and questions whose answers sit in the records or in the chain. There are no image files and no references to pixels: the dataset teaches a small language model how an artist or a photographer reasons about a picture, from a glance to a full compositional reading, so that the reasoning can later be attached to a picture by other means.

The rows are written by a language model (Claude Sonnet 5.5) from sampled specifications or from seeds, and every row passes a validator. The vocabulary, the specification sampler and the validator ship with the dataset in tools/thought_tools.py (Python, numpy only; see tools/README.md); the seed importers are the thought-image-seeds package (github.com/AbstractEyes/thought-image-seeds).

Status: pilot release

part	now	where
rows	1,995 (simple 903, medium 679, deep 413; photograph 1,382, painting 368, illustration 245; 1,242 from sampled specifications, 753 from seeds)	data/ (the default config)
seeds	1,116,536 from eleven sources (18,155 of them held; the table below)	seeds/ (the seeds config)

The pilot is complete: its rows are all here, and the next volume follows a read of the pilot by eye. The pilot's seeded rows come from the first two sources below; the other nine (scene graphs, a described-image set, museum open-access collections, Wikidata's artworks, prompt sets) join the next volume. The held split (rows on the sealed values) comes with the reads it serves.

The seed sources

source	seeds	held	license	domains
COCO captions (coco)	121,382	727	CC BY 4.0	photograph 121,382
Localized Narratives (localized_narratives)	116,602	492	CC BY 4.0	photograph 116,602
GQA (Visual Genome scene graphs) (gqa)	83,093	1,618	CC BY 4.0	photograph 83,093
DOCCI (docci)	11,877	144	CC BY 4.0	photograph 11,877
The Metropolitan Museum of Art (met)	186,979	2,632	CC0 1.0	design 100,865, illustration 49,637, photograph 14,943, sculpture 11,808, painting 9,726
Art Institute of Chicago (aic)	3,946	54	CC0 1.0	design 2,218, illustration 716, painting 656, sculpture 187, photograph 169
National Gallery of Art (nga)	59,932	1,965	CC0 1.0	illustration 49,076, painting 4,156, sculpture 3,480, photograph 2,750, design 470
Cleveland Museum of Art (cleveland)	20,684	396	CC0 1.0	design 9,860, illustration 6,214, painting 2,554, sculpture 1,232, photograph 824
Wikidata (wikidata) *	15,629	149	CC0 1.0	painting 13,260, sculpture 2,369
PartiPrompts (parti)	1,331	6	Apache-2.0	illustration 1,163, painting 88, photograph 72, design 8
DiffusionDB (diffusiondb)	495,081	9,972	CC0 1.0	illustration 350,981, photograph 74,741, painting 66,028, design 3,331

* a pictures.json beside the shards names the seeds whose picture another source holds too (a Wikidata item and the museum seed of the same artwork); a picture is drawn into one row only.

The three depths

depth	the thought holds	records	thoughts	questions	wordings
simple	one observation or one naming	1 to 4	1 to 2	1 to 2	1
medium	subjects with their attributes, their relations, the form, a definition, the behaviour shown	6 to 14	3 to 5	3 to 5	1
deep	the whole composition, mood and intent with their evidence, a comparison or a change, the story, a critique	14 to 30	6 to 10	5 to 8, one of them unanswerable from the records and one answered only by the chain	3 (caption, prompt, tags)

Rows come in series of three on one hidden image: ranks 1, 2 and 3 are the simple, medium and deep readings of the same specification, each specification a subset of the next.

The specification: a trie of systems

A record is system attribute: value. The systems and their attributes:

system	attributes
person	count, gender, age, role
look	hair, eyes, expression, makeup, jewelry, skin
wear	clothing, accessory, colour
body	pose, form, gesture
scene	setting, time, weather, layout, mood
shot	light, angle, framing, lens, focus
thing	name, count, colour, state, position
form	compose, balance, line, shape, space, rhythm, contrast, palette, texture, medium, style, period
mood	feeling, tone, energy, temperature, from (the key of the record the mood is read from)
intent	purpose, for (the audience), message, genre
story	moment, before, after, cause, teller
relation	a list of {a, rel, b} with rel from a closed list (left of, above, behind, holding, facing, larger than, ...)

Values are plain lower-case words of at most 20 bytes, and a key with its value is at most 27 bytes, so every record fits one 32-byte slice of the frame grammar the rows are trained in. Forty-one attributes draw from closed lists (vocab/trie.json); the rest are open. A tenth of every closed list is sealed (vocab/held.json): training rows never use a sealed value, and the held split's rows use at least one, so a model trained on train can be read on words it never saw in a record.

The schema

One JSON object per line:

field	content
id, series, rank, depth, domain, genre	the row's identity: art-NNNNNN, s-NNNNN, 1 to 3, simple / medium / deep, photograph / painting / illustration / sculpture / design, a free genre word
spec	the hidden image: {system: {attribute: value}} plus relation: [{a, rel, b}]
thoughts	a list of {step, text}; step is one of see, name, place, count, relate, form, light, colour, mood, mean, intend, story, compare, change, check, say
wordings	{caption, prompt, tags, critique}; the prompt is the phrase list that would recreate the picture; tags use the booru register where it exists
questions	a list of {key, value, needs}; needs is record (the value is a record of the spec), chain (the value is stated by the thoughts under a key the spec lacks) or unknown (a key the spec does not hold; the value is the word unknown)
held	{layout, wordings, values}: what this row holds back for the reads (a frame layout name, whether its wordings are held, the sealed values it uses)
source	{kind, model, seed_dataset, license}

An example row of depth medium:

{"id": "art-000123", "series": "s-00041", "depth": "medium", "rank": 2, "domain": "photograph", "genre": "street portrait",
 "spec": {"person": {"gender": "woman"}, "look": {"hair": "short", "expression": "wry"},
          "wear": {"clothing": "raincoat", "colour": "yellow"},
          "scene": {"setting": "market street", "time": "dusk", "weather": "rain"},
          "shot": {"light": "neon", "framing": "waist up"}, "form": {"compose": "right third"},
          "relation": [{"a": "woman", "rel": "in front of", "b": "fish stall"}],
          "mood": {"feeling": "amused", "from": "look expression"}},
 "thoughts": [{"step": "see", "text": "A woman in a yellow raincoat, short hair, a wry half smile."},
              {"step": "place", "text": "She stands on a market street at dusk, in the rain, in front of a fish stall."},
              {"step": "form", "text": "Neon lights her face; she sits in the right third of the frame, seen from the waist up."},
              {"step": "mean", "text": "Wry reads as amused. Dusk and neon leave a single subject lit: an inference."},
              {"step": "say", "text": "The composed wording follows."}],
 "wordings": {"caption": "A woman with short hair and a wry smile, in a yellow raincoat, on a rainy market street at dusk, lit by neon, in front of a fish stall.",
              "prompt": "street portrait, woman, short hair, wry smile, yellow raincoat, rainy market street at dusk, neon light, fish stall behind, waist up, subject in the right third",
              "tags": "1girl, short hair, smile, raincoat, rain, dusk, neon lights, market, portrait, upper body"},
 "questions": [{"key": "look hair", "value": "short", "needs": "record"},
               {"key": "rel1 b", "value": "fish stall", "needs": "record"},
               {"key": "scene layout", "value": "single subject", "needs": "chain"},
               {"key": "wear accessory", "value": "unknown", "needs": "unknown"}],
 "held": {"layout": null, "wordings": false, "values": []},
 "source": {"kind": "synthetic", "model": "sonnet", "seed_dataset": null, "license": "Apache-2.0"}}

A row written from a seed (a specification drawn from a licensed source dataset) carries source = {"kind": "seeded", "model": ..., "seed_dataset": <dataset>, "seed_id": <source>:<id>, "license": <the seed's license>}, so every represented fact keeps its attribution.

The rules every row passes

1. Every claim is in the specification or is marked as an inference with its evidence; nothing refers to pixels or to seeing a file (the medium pixel art is a medium, not a file).
2. Record values are plain lower-case words of at most 20 bytes under keys from the trie, key and value together at most 27 bytes; longer language lives in the thoughts and the wordings.
3. The depth sets the counts of records, thoughts and questions.
4. Every question's value is a record's value, the word unknown under a key the specification does not hold, or a value the thoughts state under a key the specification lacks.
5. At least one wording a row, three for deep (caption, prompt, tags); tags use the booru register where it exists.
6. A mood names what it is read from: mood from holds the key of the record that shows it.
7. A series shares its specification by subset: simple within medium within deep.
8. No real person's name, and no copyrighted image described from memory: every specification is invented or comes from a licensed description.

Layout

data/<depth>/<domain>/<domain>-<depth>-NNNNN.jsonl   the training rows (10,000 a shard)
held/<shard>.jsonl                                   the rows kept for the reads (sealed values; whole series, never split); to come
seeds/<source>/<source>-NNNNN.jsonl                  the seeds: specifications drawn from licensed source datasets, with provenance
seeds/<source>/stats.json                            per source: the counts by domain, genre and record, the held seeds
seeds/<source>/pictures.json                         where present: seed id -> the picture another source holds too
seeds/README.md                                      one line per source: the dataset, its license, what was taken
vocab/trie.json                                      the systems, the keys, the closed lists, the counts, the rules
vocab/held.json                                      the sealed values (one draw from one seed)
tools/                                               the vocabulary, the sampler, the validator, the prompt (and their tests)
GENERATION.md                                        how rows are written: the depths, the rules, the schema, the seed contract
LICENSES.md, CITATIONS.bib                           every source with its license and attribution; the BibTeX of every source

Sources, attribution and license

This is a reusable public dataset, released under Apache-2.0. The rows are written by a language model, from specifications it invents or from seeds: specifications drawn from public datasets whose licenses are permissive (CC0, CC BY, ODC-By, Apache-2.0, MIT). Every such source is attributed and cited in LICENSES.md, and every seeded row names its source and its license in its source field. Share-alike and non-commercial sources are not used. No image from any source is part of this dataset.

Citations

The sources' citations (BibTeX) are in CITATIONS.bib, and LICENSES.md lists every source with its license and attribution; cite them with this dataset when you use rows or seeds derived from them. Rows written from a CC BY or ODC-By seed keep the seed's license in their source field, which carries the attribution with the row.

Citation

@dataset{thought_symbolic_image_captions_2026,
  title  = {Thought-Symbolic Image Captions: text-only chains of thought about images at three depths},
  author = {AbstractPhil},
  year   = {2026},
  url    = {https://huggingface.co/datasets/AbstractPhil/thought-symbolic-image-captions}
}