Koa-Chang/TissueMNIST-224-full-gpt5nano-with-vlm-features

TissueMNIST 224 Full Train Val with GPT-5-nano VLM Features

The full TissueMNIST train and validation splits with categorical morphology features generated by GPT-5-nano. Test is included as the full TissueMNIST passthrough split with null vlm_model_name and placeholder vlm_feature values for schema consistency.

This dataset is derived from the official MedMNIST TissueMNIST 224px data. The VLM feature labels are categorical privileged-information annotations for CS231N VLM-LUPI experiments, not clinical diagnoses.

Splits

Columns

• —image: raw 224x224 uint8 grayscale TissueMNIST image.
• —label: TissueMNIST class label.
• —dataset_index: official index within the source MedMNIST split.
• —vlm_model_name: model that generated the categorical VLM features, or null for passthrough splits.
• —vlm_feature: integer-coded categorical morphology features in the order below. Passthrough splits contain zero placeholders and are marked in metadata.json.

Indexing

Use (split, dataset_index) as the stable MedMNIST identity. dataset_index is the zero-based index within the official MedMNIST TissueMNIST 224px source split; it is not a newly assigned subset-local identifier.

The train and validation splits use the same source indices as the official MedMNIST arrays. For example, train row with dataset_index = 42 here and in Koa-Chang/TissueMNIST-224-1pct-gpt55-with-vlm-features refer to the same official MedMNIST train[42] datapoint when that index appears in both repos.

dataset_index values repeat across source splits, so keep the split name when joining or comparing rows. If you need one flat integer across TissueMNIST 224, use global_index = split_offset[split] + dataset_index, with split offsets train=0, val=165466, and test=189106.

VLM Feature Order

• —background_clutter: cleandark, hazybackground, brightneighbors, noisyartifact
• —dominant_region_scale: small, medium, large
• —dominant_shape: roundoval, elongatedspindle, linearstrand, irregularblob, fragmented_cluster
• —edge_contact: none, minor, major
• —elongation: low, moderate, high
• —foreground_fraction: sparse, moderate, dense
• —internal_signal_texture: smooth, finespeckled, coarsemottled, brightpuncta, rimor_halo
• —local_arrangement: isolated, pairedoradjacent, crowdedcluster, linearchain, sheetortubule_like, overlapping
• —raw_contrast: low, moderate, high
• —signal_distribution: uniform, centralenriched, peripheralrim, multifocal, diffuse_patchy
• —visible_object_count: none, single, twotothree, fourormore, unclear

Source Label Directories

• —shared/vlm_labels/tissuemnist_train_gpt5nano_full_v2
• —shared/vlm_labels/tissuemnist_val_gpt5nano_full_v2

Loading

from datasets import load_dataset

ds = load_dataset("Koa-Chang/TissueMNIST-224-full-gpt5nano-with-vlm-features")
print(ds)

The project training code can load this repo through data.huggingface.from_huggingface("Koa-Chang/TissueMNIST-224-full-gpt5nano-with-vlm-features").