Remove dataset card

README.md

@@ -1,120 +0,0 @@
----
-pretty_name: MoRA Missing-Modality Benchmark Data
-license: other
-language:
-- en
-task_categories:
-- image-classification
-- text-classification
-tags:
-- multimodal
-- missing-modality
-- vision-language
-- clip
-- mm-imdb
-- food101
-- hateful-memes
-size_categories:
-- 10K<n<100K
-viewer: false
----
-
-# MoRA — Missing-Modality Benchmark Data
-
-Preprocessed, ready-to-train data for reproducing **MoRA: Missing Modality Low-Rank Adaptation for Visual Recognition**.
-
-- 📄 Paper: [arXiv:2511.06225](https://arxiv.org/abs/2511.06225)
-- 💻 Code: [github.com/Tree-Shu-Zhao/MoRA](https://github.com/Tree-Shu-Zhao/MoRA)
-
-This bundle exists so you don't have to redo the (error-prone) raw download + preprocessing. The Arrow files and JSON metadata drop **directly** into the repo's `datasets/` directory with no renaming — filenames already match what `src/data/dataset.py` expects.
-
-## Quick start
-
-```bash
-# from the MoRA repo root
-huggingface-cli download TreezzZ/MoRA-data --repo-type dataset --local-dir datasets
-# then, per the README:
-python src/main.py experiment=mora_mmimdb
-python src/main.py experiment=mora_hatememes
-python src/main.py experiment=mora_food101
-```
-
-That's it — `DATA_DIR` defaults to `datasets`, so the three dataset folders line up automatically.
-
-## Directory structure
-
-```
-.
-├── Food101/
-│   ├── train.arrow            # image bytes + text + label
-│   ├── val.arrow
-│   ├── test.arrow
-│   ├── class_idx.json         # 101 class → index map
-│   ├── text.json              # per-sample UPMC text titles
-│   ├── split.json             # train/val/test sample ids
-│   └── missing_tables/        # precomputed missing-modality masks (see below)
-├── hateful_memes/
-│   ├── train.arrow
-│   ├── val.arrow              # = dev split
-│   ├── test.arrow             # labelled test (from test_seen/test_unseen)
-│   ├── test_unseen.jsonl
-│   └── missing_tables/
-└── mmimdb/
-    ├── train.arrow
-    ├── val.arrow              # = dev split
-    ├── test.arrow
-    ├── split.json
-    └── missing_tables/
-```
-
-### Arrow schema
-Each `*.arrow` is a PyArrow IPC file. Columns:
-- **MM-IMDb**: `image` (JPEG bytes), `plots` (list[str]), `label` (23-d multi-hot), `genres`, `image_id`, `split`
-- **Hateful Memes**: `image` (PNG bytes), `text` (list[str]), `label` (int), `split`
-- **Food-101**: `image` (JPEG bytes), `text` (list[str]), `label` (int), `image_id`, `split`
-
-### `missing_tables/`
-Precomputed missing-modality masks used by the paper. Naming convention (matches `src/data/dataset.py`):
-
-```
-{dataset}_split{split}_missing_{type}_{ratio}.pt
-  dataset ∈ {food101, hatememes, mmimdb}
-  split   ∈ {train, val, test}
-  type    ∈ {both, image, text}
-  ratio   ∈ {0.1, 0.2, …, 0.9}
-```
-
-Each file is a 1-D `torch` tensor of length *N* (number of samples): `0` = both modalities present, `1` = text missing, `2` = image missing. When a file matching the active `missing_params` exists at `datasets/<DS>/missing_tables/`, the dataloader **loads it**; otherwise it regenerates one deterministically from `SEED`. Shipping these guarantees everyone evaluates on identical masks.
-
-## Reproduction
-
-The repo's default config is **missing type `both`, ratio `0.7`** (`BOTH_RATIO=0.5`). Using this bundle's masks, a single H100 run per dataset reproduces the paper's Table 1 (70%, "both") within run-to-run noise:
-
-| Dataset | Metric | Paper | Reproduced |
-|---|---|---|---|
-| MM-IMDb | F1-macro | 52.97 | 52.53 |
-| Hateful Memes | AUROC | 70.15 | 69.56 |
-| Food-101 | Accuracy | 83.77 | 83.30 |
-
-(Residual ≤0.6 pt is fp16/cuDNN training nondeterminism; masks are identical.)
-
-## Data provenance & licensing
-
-This is a **research convenience redistribution** of three public benchmarks. The original data are owned by their respective creators and remain under their original terms — **you are responsible for complying with each source's license/usage agreement**:
-
-- **MM-IMDb** — [gmu-mmimdb](https://github.com/johnarevalo/gmu-mmimdb) / [archive.org mirror](https://archive.org/download/mmimdb)
-- **UPMC Food-101** — [VISIIR](https://visiir.isir.upmc.fr/explore) / [Kaggle](https://www.kaggle.com/datasets/gianmarco96/upmcfood101)
-- **Hateful Memes** — [Meta AI](https://ai.facebook.com/blog/hateful-memes-challenge-and-data-set/) (subject to the Hateful Memes Dataset License Agreement; redistribution of the original images may be restricted — use accordingly)
-
-If you are a rights holder and have concerns about redistribution, please contact the repository owner.
-
-## Citation
-
-```bibtex
-@article{zhao2025mora,
-  title={MoRA: Missing Modality Low-Rank Adaptation for Visual Recognition},
-  author={Zhao, Shu and Ahuja, Nilesh and Yu, Tan and Shen, Tianyi and Narayanan, Vijaykrishnan},
-  journal={arXiv preprint arXiv:2511.06225},
-  year={2025}
-}
-```