xzz789/novel-entity-100

Novel Entity 100

This dataset contains 100 original procedurally modeled abstract objects, each with a unique nonce name and a fixed, entirely fictional maker record. It provides 10,000 RGB images at 512×512, one portable GLB mesh and one preview per object, and paired data for visual understanding and text-to-image generation. The objects and all maker records are newly constructed for this dataset; no CLEAR identities, biographies, images, or external meshes are included.

Each object has 60 training views, 20 held-out camera views, and 20 held-out appearance views with changed background and lighting. A held-out appearance view may share a camera direction with training. The same geometric object appears in all three groups; this is an identity-generalization split, not a disjoint-object split.

The maker's name, profession, hometown, education, and workshop, plus the object's completion year, series, stated inspiration, and first exhibition are assigned synthetic facts. They cannot be inferred from the image without learning this specific model–record binding. The abstract object is not personified: it has no birth date or parents. All maker names, venues, and workshops are fictional.

Load the images and metadata

from datasets import load_dataset

ds = load_dataset("xzz789/novel-entity-100", data_dir="images")
assert len(ds["train"]) == 6000 and len(ds["test"]) == 4000
sample = ds["train"][0]
image = sample["image"]              # decoded PIL image
question, answer = sample["question"], sample["answer"]
name_prompt = sample["caption"]      # name-conditioned T2I prompt

The test split has eval_group values unseen_view and unseen_appearance (2,000 rows each). images/{train,test}/metadata.jsonl links each image to its metadata through file_name; source_image is the repository-relative file path. The caption column remains name-only for generation. biography is the complete maker-and-object description for a separately chosen caption task. It is not an implicit part of the name-conditioned generation task.

Files for other training recipes

annotations/ contains the fixed 100-record catalog (profiles.jsonl), the profile schema, and paired train/test JSONL exports for image-field QA, full-record captioning, name-conditioned generation, and text-only QA. train_clear_mixed.jsonl and test_clear_mixed.jsonl already combine the caption and text-only rows; do not concatenate them again with their component files. Each annotation image path is relative to this dataset repository. annotations/generation_profile_queries.jsonl has 100 name-hidden maker-fact prompts reserved for additional evaluation, not name-prompt training.

models/ contains 100 original GLB meshes, keyed by the same IDs as the images. previews/ contains one render per mesh; annotations/models.jsonl maps model IDs, names, structural families, GLB paths, hashes, and maker profile hashes. Native Blender project files are not distributed because they contain local authoring-environment paths; the GLB geometry is portable and sufficient to inspect or render each model. SOURCE.json and SHA256SUMS.jsonl record source and release fingerprints.

Intended use and limits

The data supports controlled studies of whether a unified multimodal model learns, transfers, and forgets the same object identity and its assigned maker facts across understanding and generation. For forgetting an ID, filter that ID consistently across every selected training stream and evaluate both paths on held-out views. The data does not establish that any base model or target-learned checkpoint already knows these objects. It is not a dataset of human faces or real-person identities. Maker-fact-conditioned generation is an evaluation proposal, not a validated benchmark score.

The model roster, geometry, and image bytes are fixed by the source fingerprint in SOURCE.json. Rendering and fact assignment are synthetic; results on these abstract objects may not transfer to human-identity unlearning. The construction is inspired by the idea of fixed multimodal identity records in CLEAR, but this release is an independent dataset.

License

No reuse license is declared.