Correct limitations; link training datasets

README.md

@@ -7,6 +7,19 @@ tags:
   - pytorch
   - flow-matching
   - art
+datasets:
+  - kaupane/chinese-painting-collection
+  - kaupane/wikiart-captions
+  - kaupane/human-recaption
+  - kaupane/pexels-people-captions
+  - kaupane/relaion-art-recap-zh
+  - kaupane/vintage-photography-captions
+  - kaupane/z-image-turbo-gen
+  - kaupane/inko-rollout-25k
+  - opendiffusionai/cc12m-a_woman
+  - Spawning/PD12M
+  - sirus/inaturalist-2024-2.8k-claude-opus-5-recaptioned
+  - sirus/megalith-10m-5.5k-claude-opus-5-recaptioned
 ---
 
 <h1 align="center">Inko</h1>
@@ -82,9 +95,19 @@ The release defaults are 896 × 896, 20 Euler steps, and CFG 4.5. Width and heig
 - [Post-training](https://github.com/Mtrya/inko/blob/main/notes/posttraining.md): the DiffusionART selection campaign, the 24.6k-row rollout pool ([kaupane/inko-rollout-25k](https://huggingface.co/datasets/kaupane/inko-rollout-25k)), the rubric reward, and the two production runs.
 - [Evaluation](https://github.com/Mtrya/inko/blob/main/notes/evaluation.md): five benchmark catalogs, nine models, and a fixed fine-art comparison suite with per-case analysis.
 
+## Training data
+
+The pretraining corpus combines museum collections, recaptioned public datasets and synthetic additions across four domains. The sources available on the Hub:
+
+- Chinese painting: [kaupane/chinese-painting-collection](https://huggingface.co/datasets/kaupane/chinese-painting-collection)
+- Western art: [kaupane/wikiart-captions](https://huggingface.co/datasets/kaupane/wikiart-captions)
+- People and anatomy: [kaupane/human-recaption](https://huggingface.co/datasets/kaupane/human-recaption) (a recaption of [OpenFace-CQUPT/HumanCaption-HQ-311K](https://huggingface.co/datasets/OpenFace-CQUPT/HumanCaption-HQ-311K)), [opendiffusionai/cc12m-a_woman](https://huggingface.co/datasets/opendiffusionai/cc12m-a_woman), [kaupane/pexels-people-captions](https://huggingface.co/datasets/kaupane/pexels-people-captions)
+- World knowledge: [kaupane/relaion-art-recap-zh](https://huggingface.co/datasets/kaupane/relaion-art-recap-zh), [Spawning/PD12M](https://huggingface.co/datasets/Spawning/PD12M), [sirus/inaturalist-2024-2.8k-claude-opus-5-recaptioned](https://huggingface.co/datasets/sirus/inaturalist-2024-2.8k-claude-opus-5-recaptioned), [sirus/megalith-10m-5.5k-claude-opus-5-recaptioned](https://huggingface.co/datasets/sirus/megalith-10m-5.5k-claude-opus-5-recaptioned), [kaupane/vintage-photography-captions](https://huggingface.co/datasets/kaupane/vintage-photography-captions), [kaupane/z-image-turbo-gen](https://huggingface.co/datasets/kaupane/z-image-turbo-gen)
+- Post-training rollout pool: [kaupane/inko-rollout-25k](https://huggingface.co/datasets/kaupane/inko-rollout-25k)
+
 ## Limitations
 
-- Exact rendered text, detector-based counting and spatial relations remain weak.
-- Some unrelated subjects can be substituted by ornate portraits; requested illustration styles and actions can collapse into static realistic portraits.
+- Exact rendered text and spatial relations remain weak.
+- Anatomy still breaks at times; reasoning, counting and world knowledge remain weak.
 - Negation in dense scenes is unreliable: asked for a busy intersection with no blue cars, Inko, like most references, paints the cars.
 - The corpus includes research-only sources; per-sample license metadata is preserved in the training configuration.