Document source dataset licenses and retain mixed licensing

LICENSE.md

@@ -0,0 +1,106 @@
+# Dataset Licensing
+
+OmniTaskonomy contains data from multiple sources with different licenses.
+Each component remains subject to its original license. No relicensing of
+third-party content is intended.
+
+The sections below describe the source datasets used by OmniTaskonomy.
+[release_manifest.json](release_manifest.json) lists the splits currently available.
+Each sample records its source in `source_dataset`, `source_id`, and
+`source_metadata`. A benchmark's license declaration does not replace the
+terms governing its underlying images.
+
+## CC BY 4.0
+
+- **MMT-Bench**: the benchmark is released under CC BY 4.0.
+  [Source and license](https://huggingface.co/datasets/OpenGVLab/MMT-Bench).
+- **COCO annotations**: used for semantic segmentation and COCO-based tasks.
+  Image rights are separate; see below.
+  [Official terms](https://cocodataset.org/#termsofuse).
+- **LVIS annotations**: used by VisGym object pointing; images come from COCO.
+  [Authors' license clarification](https://groups.google.com/g/lvis-dataset/c/JeWfhD3vioM).
+- **Google RefExp / RefCOCOg annotations**: used for localization, obtained from
+  [jxu124/refcocog](https://huggingface.co/datasets/jxu124/refcocog).
+  [Original release and license](https://github.com/mjhucla/Google_Refexp_toolbox#license).
+
+Users must comply with the original
+[CC BY 4.0 license](https://creativecommons.org/licenses/by/4.0/)
+and attribution requirements.
+
+## CC BY-ND 4.0
+
+**RealWorldQA**: this dataset contains a sampled subset of
+[xai-org/RealworldQA](https://huggingface.co/datasets/xai-org/RealworldQA).
+The original samples remain subject to
+[CC BY-ND 4.0](https://creativecommons.org/licenses/by-nd/4.0/).
+
+## Apache 2.0 declarations
+
+The following upstream dataset cards declare Apache 2.0:
+
+- **BLINK**: [BLINK-Benchmark/BLINK](https://huggingface.co/datasets/BLINK-Benchmark/BLINK).
+- **CV-Bench**, including its 2D and 3D subsets:
+  [nyu-visionx/CV-Bench](https://huggingface.co/datasets/nyu-visionx/CV-Bench).
+- **ConceptEdit-12M**: used for object editing, attribute editing, and part of
+  inpainting. [inclusionAI/ConceptEdit-12M](https://huggingface.co/datasets/inclusionAI/ConceptEdit-12M).
+  Its image source, [Fine-T2I](https://huggingface.co/datasets/ma-xu/fine-t2i),
+  also declares Apache 2.0.
+- **VisGym training data**: used for jigsaw and object pointing.
+  [VisGym/visgym_data](https://huggingface.co/datasets/VisGym/visgym_data).
+
+The [Apache 2.0 license](https://www.apache.org/licenses/LICENSE-2.0)
+applies within the scope of each upstream release. Third-party images retain
+their original terms. BLINK and CV-Bench draw on other datasets; Fine-T2I
+includes generated images and photographs from external providers.
+
+## MIT declaration
+
+**MMVP** declares MIT in its
+[dataset card](https://huggingface.co/datasets/MMVP/MMVP).
+Its images originate from ImageNet and LAION-Aesthetics and retain their
+underlying image rights. The dataset-level declaration does not grant new
+rights to those images.
+
+## Taskonomy data terms
+
+**Taskonomy**, used for colorization, depth, normals, curvature, edges,
+keypoints, reshading, and unsupervised segmentation, has a
+[custom dataset license](https://github.com/StanfordVL/taskonomy/blob/master/data/LICENSE).
+It limits use to non-commercial or not-for-profit research or education,
+and requires prior written permission for redistribution. The MIT license
+for Taskonomy's code is separate from its data license. The
+[Omnidata license](https://github.com/EPFL-VILAB/omnidata/blob/main/LICENSE)
+also contains data-specific terms and incorporates the Taskonomy agreement.
+
+The 12 Taskonomy-derived splits are not currently published; written
+redistribution permission is being requested.
+
+## Underlying image rights and unspecified terms
+
+- **COCO images**, used in inpainting, semantic segmentation, localization,
+  and LVIS object pointing, retain their individual image licenses.
+  COCO's CC BY 4.0 annotation license does not cover all images.
+  [COCO image terms](https://cocodataset.org/#termsofuse).
+- **CV-Bench images** come from [COCO](https://cocodataset.org/#termsofuse), [ADE20K](https://ade20k.csail.mit.edu/terms/) (non-commercial research and education), and [Omni3D source datasets](https://github.com/facebookresearch/omni3d/blob/main/DATA.md):
+  [Hypersim](https://github.com/apple-aiml-research/ml-hypersim) (CC BY-SA 3.0), [SUN RGB-D](https://rgbd.cs.princeton.edu/), and [nuScenes](https://www.nuscenes.org/terms-of-use).
+  These source-image terms are separate from CV-Bench's Apache 2.0 declaration.
+- **VisGym jigsaw images** retain their original image terms. VisGym's
+  [default image download](https://github.com/visgym/VisGym/blob/main/download_data.sh)
+  uses [LLaVA-Pretrain](https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain),
+  whose image release is intended for research reproduction and refers to
+  the original LAION/CC/SBU terms. Episode metadata does not establish a
+  separate license for every image.
+- **V*Bench / VStarBench** uses images from SA-1B, according to its
+  [paper](https://arxiv.org/abs/2312.14135). These images are subject to the
+  [SA-1B Dataset Research License](https://ai.meta.com/datasets/segment-anything/).
+  A separate license for the benchmark annotations was not identified in
+  the upstream release. The project's MIT code license is not an image license.
+- **MMStar**: no explicit dataset license was identified in its
+  [official dataset card](https://huggingface.co/datasets/Lin-Chen/MMStar)
+  or [repository](https://github.com/MMStar-Benchmark/MMStar).
+  It incorporates samples from other benchmarks, recorded in source metadata;
+  those components retain their original terms.
+
+Where the upstream release does not specify a data license, this document
+does not supply one. Users must comply with the applicable original terms,
+including attribution, research-use restrictions, and image-specific rights.

README.md

@@ -112,12 +112,6 @@ the templates remain available in source metadata.
 
 ## License
 
-This dataset contains a sampled subset of RealWorldQA.
-The original RealWorldQA data is licensed under CC BY-ND 4.0.
-
-Source: xai-org/RealworldQA
-
-The original RealWorldQA samples remain subject to the
-CC BY-ND 4.0 license.
-
-Other source datasets retain their original terms; source information is recorded with each sample.
+This dataset combines sources with different licenses. See [LICENSE.md](LICENSE.md)
+for each source's terms, including RealWorldQA's CC BY-ND 4.0 license.
+Each component retains its original license; no relicensing of third-party content is intended.