manifest v1: the picture list (21 sources, every group; the pilot split of 120 a source), its parquet twin, counts and the card

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+manifest_v1.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,132 @@
+---
+license: other
+license_name: per-source
+license_link: https://huggingface.co/datasets/AbstractPhil/qwen-thinking-images#licences
+task_categories:
+- image-to-text
+language:
+- en
+pretty_name: qwen-thinking-images
+size_categories:
+- 10K<n<100K
+tags:
+- manifest
+- image-captioning
+- chain-of-thought
+- symbolic-thinking
+configs:
+- config_name: default
+  data_files: manifest_v1.parquet
+---
+
+# qwen-thinking-images
+
+The list of pictures the symbolic-thinking VLM ([AbstractPhil/qwen3.5-9b-think-symbolic-image](https://huggingface.co/AbstractPhil/qwen3.5-9b-think-symbolic-image),
+a LoRA on Qwen3.5-9B) writes whole rows about: thoughts, wordings, and the questions a careful looker would still ask. This
+repository holds **the list, not the pictures**: every row says where its picture lives (a public url, a row of a parquet shard,
+a file or a zip member on the Hub) and what its source already says about it, so the same set can be drawn again, extended, and
+compared across model versions. The model's rows ship here too, under `extractions/`.
+
+The set was drawn for breadth rather than size: a couple of thousand pictures from each major source, photographs and art and
+synthetic renders alike, portraits and clothes (FFHQ, DeepFashion), rendered single words for word reading, social posts with
+their poor captions, a light slice of anime renders, and the synthetic ImageNet set spread over its classes (at most four
+pictures a class, 998 classes). Pictures that already had a human-written thought row in
+[thought-symbolic-image-captions](https://huggingface.co/datasets/AbstractPhil/thought-symbolic-image-captions), or that served as
+synthetic seeds or screen items for the model, are excluded, so nothing here was seen in training.
+
+| | |
+|---|---|
+| rows | 48,145 |
+| pilot split | 2,520 (120 a source) |
+| groups | photo 12,978, synthetic 12,000, art 9,369, portrait 4,500, fashion 2,500, social 2,000, render3d 2,000, text 2,000, anime 798 |
+| depth plan | quick 15,806 / closer 22,880 / long 9,459 |
+
+## Composition
+
+| source | group | rows | pilot | captions | synthetic | pictures live at | licence |
+|---|---|---:|---:|---:|---|---|---|
+| big_liminal | photo | 978 | 120 | 978 | no | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| coco | photo | 6,000 | 120 | 5,996 | no | cocodataset.org (url) | CC BY 4.0 |
+| docci | photo | 3,000 | 120 | 2,994 | no | google/docci (hf_dataset_row) | CC BY 4.0 |
+| localized_narratives | photo | 3,000 | 120 | 2,977 | no | cocodataset.org (Localized Narratives) (url) | CC BY 4.0 |
+| aic | art | 1,990 | 120 | 1,947 | no | artic.edu (url) | CC0 1.0 |
+| cleveland | art | 2,932 | 120 | 2,932 | no | clevelandart.org (url) | CC0 1.0 |
+| nga | art | 3,000 | 120 | 3,000 | no | nga.gov (url) | CC0 1.0 |
+| wikidata | art | 1,447 | 120 | 1,447 | no | wikidata.org (url) | CC0 1.0 |
+| ffhq | portrait | 2,500 | 120 | 2,500 | no | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| imdb | portrait | 2,000 | 120 | 2,000 | no | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| deepfashion | fashion | 2,500 | 120 | 2,500 | no | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| instagram | social | 2,000 | 120 | 2,000 | no | AbstractPhil/kkcosmos_instagram-images-with-bad-captions-sd-scripts (zip_member) | see the source repository |
+| flux_assorted_bulk | synthetic | 2,000 | 120 | 2,000 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| flux_assorted_bulk_2 | synthetic | 2,000 | 120 | 2,000 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| imagenet_synthetic | synthetic | 3,000 | 120 | 3,000 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| sdxl_qwen_phase0 | synthetic | 3,000 | 120 | 3,000 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| synth_chars | synthetic | 2,000 | 120 | 2,000 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| mannequins_v10 | render3d | 1,500 | 120 | 1,500 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| mannequins_v7 | render3d | 500 | 120 | 500 | yes | AbstractPhil/diffusion-pretrain-set-ft1 (hf_parquet) | see the pretrain set's per-source licences |
+| anima | anime | 798 | 120 | 798 | yes | AbstractPhil/geolip-beatrix-anima-data (hf_file) | anima-outputs (see the source repository) |
+| trdg_words | text | 2,000 | 120 | 0 | yes | AbstractPhil/priyank-m_trdg_random_single_words_en_text_recognition-sd-scripts (zip_member) | see the source repository |
+| **all** | | **48,145** | **2,520** | | | | |
+
+"captions" counts the rows whose source carries at least one plain-text caption (the seed hint of a thought seed, a source
+caption, a drawing prompt); those rows can be audited for grounding, the rest for form only. Museum rows whose object had
+no open-access picture were dropped (Cleveland 68, the Art Institute
+10, Wikidata 553 without a P18 image).
+Two sources drawn for this list are held back from v1 and will join a later version: the Met (its collection API throttled the
+picture lookup; 3,000 objects drawn) and GQA / Visual Genome (the Stanford image host no longer answers; 2,000 drawn).
+
+## Files
+
+| file | what |
+|---|---|
+| `manifest_v1.jsonl` | the whole list, one JSON object a row (nested fields as objects) |
+| `manifest_v1.parquet` | the same rows; `image`, `captions`, `labels` and `spec` are JSON strings |
+| `pilot_v1.jsonl` | the pilot split alone: 2,520 rows, the first extraction run |
+| `counts.json` | the counts above by group, source, depth and split |
+| `extractions/<stamp>_<tag>/rows.jsonl` | the model's rows for a run (text, parsed thoughts / wordings / questions, scores, the audit's verdict) and its `summary.json` |
+
+## Fields
+
+| field | meaning |
+|---|---|
+| `sample_id` | `qti-NNNNNN`, stable across versions of this list |
+| `split` | `pilot` or `main` |
+| `group` | photo, art, portrait, fashion, social, synthetic, render3d, anime, text |
+| `source`, `source_repo`, `subset`, `source_id` | where the picture comes from and its id there |
+| `image` | the locator (below) |
+| `width`, `height` | when the source states them |
+| `captions` | what the source already says: `seed_hint` (a thought seed's hint), `source_caption`, `draw_prompt`, the pretrain set's caption kinds, `vlm_0.8b_json`, `animetimm_json` |
+| `labels` | class labels where the source has them (ImageNet `true_label`, `semantic_class`; a seed's `domain`, `genre`) |
+| `spec` | a thought seed's specification (the records a drawn row would be written from), empty otherwise |
+| `synthetic` | whether the picture is generated |
+| `license` | the source's licence statement (see below) |
+| `depth_plan` | the depth the model is asked at: `quick` (one or two thoughts, a caption), `closer` (three to five, a caption and a prompt or tags), `long` (six to ten, all four wordings) |
+
+## Locators
+
+| kind | fields | example |
+|---|---|---|
+| `url` | `urls` (tried in order) | COCO, Visual Genome, the museums (IIIF or open-access CDN), Wikimedia Commons |
+| `hf_parquet` | `repo`, `repo_type`, `path`, `row_index`, `column`, `id` | a row of a shard of the pretrain set |
+| `hf_file` | `repo`, `repo_type`, `path` | an Anima render |
+| `zip_member` | `repo`, `repo_type`, `path`, `member` | a word image or an instagram picture inside a zip |
+| `hf_dataset_row` | `repo`, `split`, `key_column`, `key`, `image_column` | a DOCCI picture by its `example_id` |
+
+`qwen_symbolic_thinking.extract.fetch_images(rows, cache_dir)` ([github](https://github.com/AbstractEyes/qwen-symbolic-thinking))
+resolves every kind, grouping the rows so a shard or a zip downloads once.
+
+## Licences
+
+The list itself (this repository's own files) is CC0. Each picture keeps the licence of its source, named in the row:
+COCO and Localized Narratives (images under their Flickr licences, annotations CC BY 4.0), Visual Genome / GQA (CC BY 4.0),
+DOCCI (CC BY 4.0), the Met, the National Gallery of Art, the Cleveland Museum of Art and the Art Institute of Chicago open-access
+pictures (CC0 for public-domain works; the row's `notes` links the object page), Wikimedia Commons (per file), the
+diffusion-pretrain-set subsets under that set's per-source terms (FFHQ, DeepFashion, IMDb faces for research use; the Flux,
+SDXL, synthetic-character, mannequin and synthetic-ImageNet renders are generated), the rendered word images and the instagram
+set under their source repositories' terms, and the Anima renders as that project's outputs. Use the pictures under their own
+terms.
+
+## Exclusions
+
+DiffusionDB and Parti prompts carry no pictures of their own and were left out; mannequin rows tagged nude and pretrain rows
+failing the set's age classifier were dropped; the Anima slice was kept light (the anime tagger already covers that ground).

counts.json

@@ -0,0 +1,340 @@
+{
+ "rows": 48145,
+ "pilot": 2520,
+ "by_group": {
+  "photo": 12978,
+  "art": 9369,
+  "portrait": 4500,
+  "fashion": 2500,
+  "social": 2000,
+  "synthetic": 12000,
+  "render3d": 2000,
+  "anime": 798,
+  "text": 2000
+ },
+ "by_source": {
+  "big_liminal": {
+   "rows": 978,
+   "group": "photo",
+   "synthetic": false,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 978,
+   "depth": {
+    "closer": 451,
+    "quick": 339,
+    "long": 188
+   },
+   "pilot": 120
+  },
+  "coco": {
+   "rows": 6000,
+   "group": "photo",
+   "synthetic": false,
+   "license": "CC BY 4.0",
+   "image_kind": "url",
+   "source_repo": "cocodataset.org",
+   "with_captions": 5996,
+   "depth": {
+    "closer": 2700,
+    "long": 1199,
+    "quick": 2101
+   },
+   "pilot": 120
+  },
+  "docci": {
+   "rows": 3000,
+   "group": "photo",
+   "synthetic": false,
+   "license": "CC BY 4.0",
+   "image_kind": "hf_dataset_row",
+   "source_repo": "google/docci",
+   "with_captions": 2994,
+   "depth": {
+    "closer": 1380,
+    "long": 620,
+    "quick": 1000
+   },
+   "pilot": 120
+  },
+  "localized_narratives": {
+   "rows": 3000,
+   "group": "photo",
+   "synthetic": false,
+   "license": "CC BY 4.0",
+   "image_kind": "url",
+   "source_repo": "cocodataset.org (Localized Narratives)",
+   "with_captions": 2977,
+   "depth": {
+    "closer": 1332,
+    "quick": 1046,
+    "long": 622
+   },
+   "pilot": 120
+  },
+  "aic": {
+   "rows": 1990,
+   "group": "art",
+   "synthetic": false,
+   "license": "CC0 1.0",
+   "image_kind": "url",
+   "source_repo": "artic.edu",
+   "with_captions": 1947,
+   "depth": {
+    "closer": 932,
+    "long": 774,
+    "quick": 284
+   },
+   "pilot": 120
+  },
+  "cleveland": {
+   "rows": 2932,
+   "group": "art",
+   "synthetic": false,
+   "license": "CC0 1.0",
+   "image_kind": "url",
+   "source_repo": "clevelandart.org",
+   "with_captions": 2932,
+   "depth": {
+    "long": 1205,
+    "closer": 1258,
+    "quick": 469
+   },
+   "pilot": 120
+  },
+  "nga": {
+   "rows": 3000,
+   "group": "art",
+   "synthetic": false,
+   "license": "CC0 1.0",
+   "image_kind": "url",
+   "source_repo": "nga.gov",
+   "with_captions": 3000,
+   "depth": {
+    "quick": 462,
+    "long": 1208,
+    "closer": 1330
+   },
+   "pilot": 120
+  },
+  "wikidata": {
+   "rows": 1447,
+   "group": "art",
+   "synthetic": false,
+   "license": "CC0 1.0",
+   "image_kind": "url",
+   "source_repo": "wikidata.org",
+   "with_captions": 1447,
+   "depth": {
+    "quick": 223,
+    "long": 591,
+    "closer": 633
+   },
+   "pilot": 120
+  },
+  "ffhq": {
+   "rows": 2500,
+   "group": "portrait",
+   "synthetic": false,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2500,
+   "depth": {
+    "closer": 1281,
+    "quick": 952,
+    "long": 267
+   },
+   "pilot": 120
+  },
+  "imdb": {
+   "rows": 2000,
+   "group": "portrait",
+   "synthetic": false,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2000,
+   "depth": {
+    "quick": 799,
+    "closer": 1000,
+    "long": 201
+   },
+   "pilot": 120
+  },
+  "deepfashion": {
+   "rows": 2500,
+   "group": "fashion",
+   "synthetic": false,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2500,
+   "depth": {
+    "quick": 725,
+    "long": 264,
+    "closer": 1511
+   },
+   "pilot": 120
+  },
+  "instagram": {
+   "rows": 2000,
+   "group": "social",
+   "synthetic": false,
+   "license": "see the source repository",
+   "image_kind": "zip_member",
+   "source_repo": "AbstractPhil/kkcosmos_instagram-images-with-bad-captions-sd-scripts",
+   "with_captions": 2000,
+   "depth": {
+    "closer": 1021,
+    "quick": 753,
+    "long": 226
+   },
+   "pilot": 120
+  },
+  "flux_assorted_bulk": {
+   "rows": 2000,
+   "group": "synthetic",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2000,
+   "depth": {
+    "closer": 1087,
+    "quick": 614,
+    "long": 299
+   },
+   "pilot": 120
+  },
+  "flux_assorted_bulk_2": {
+   "rows": 2000,
+   "group": "synthetic",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2000,
+   "depth": {
+    "closer": 1074,
+    "long": 298,
+    "quick": 628
+   },
+   "pilot": 120
+  },
+  "imagenet_synthetic": {
+   "rows": 3000,
+   "group": "synthetic",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 3000,
+   "depth": {
+    "closer": 1683,
+    "quick": 882,
+    "long": 435
+   },
+   "pilot": 120
+  },
+  "sdxl_qwen_phase0": {
+   "rows": 3000,
+   "group": "synthetic",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 3000,
+   "depth": {
+    "long": 463,
+    "closer": 1642,
+    "quick": 895
+   },
+   "pilot": 120
+  },
+  "synth_chars": {
+   "rows": 2000,
+   "group": "synthetic",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 2000,
+   "depth": {
+    "long": 325,
+    "closer": 1111,
+    "quick": 564
+   },
+   "pilot": 120
+  },
+  "mannequins_v10": {
+   "rows": 1500,
+   "group": "render3d",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 1500,
+   "depth": {
+    "closer": 736,
+    "long": 149,
+    "quick": 615
+   },
+   "pilot": 120
+  },
+  "mannequins_v7": {
+   "rows": 500,
+   "group": "render3d",
+   "synthetic": true,
+   "license": "see the pretrain set's per-source licences",
+   "image_kind": "hf_parquet",
+   "source_repo": "AbstractPhil/diffusion-pretrain-set-ft1",
+   "with_captions": 500,
+   "depth": {
+    "long": 45,
+    "closer": 245,
+    "quick": 210
+   },
+   "pilot": 120
+  },
+  "anima": {
+   "rows": 798,
+   "group": "anime",
+   "synthetic": true,
+   "license": "anima-outputs (see the source repository)",
+   "image_kind": "hf_file",
+   "source_repo": "AbstractPhil/geolip-beatrix-anima-data",
+   "with_captions": 798,
+   "depth": {
+    "closer": 473,
+    "quick": 245,
+    "long": 80
+   },
+   "pilot": 120
+  },
+  "trdg_words": {
+   "rows": 2000,
+   "group": "text",
+   "synthetic": true,
+   "license": "see the source repository",
+   "image_kind": "zip_member",
+   "source_repo": "AbstractPhil/priyank-m_trdg_random_single_words_en_text_recognition-sd-scripts",
+   "with_captions": 0,
+   "depth": {
+    "quick": 2000
+   },
+   "pilot": 120
+  }
+ },
+ "by_depth": {
+  "closer": 22880,
+  "quick": 15806,
+  "long": 9459
+ },
+ "dropped_unresolved": {
+  "aic": 10,
+  "cleveland": 68,
+  "wikidata": 553
+ }
+}

manifest_v1.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c88742e245733afc789c55df97f658e82554d88a14ae2566045f5b3953c80627
+size 59775051

manifest_v1.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4aaa66a3d24d0ef56ca265168d1f1c0e620536e98df341cca0f20b9a00ff4854
+size 6388636