Update README.md

README.md

@@ -168,3 +168,67 @@ configs:
   - split: struct_aware_parse_train
     path: data/struct_aware_parse_train-*
 ---
+### TABLET-Small
+
+This is the _Small_ sized **train set** of the **TABLET** dataset. It contains the train examples for 14 **TABLET** tasks.  
+Each task is capped at **140,000 examples**, resulting in a total of **776,602 training examples** across **14 tasks**.  
+This dataset is self-contained, each example includes a table image, its HTML representation, and the associated task data.  
+If you're interested in all tables in TABLET, see [TABLET-tables](https://huggingface.co/datasets/alonsoapp/TABLET-tables).  
+
+All TABLET Subsets:
+- _(train)_ [**TABLET-Small**](https://huggingface.co/datasets/alonsoapp/TABLET-Small): The smallest TABLET subset, including **776,602 examples** across **14 tasks**.
+- _(train)_ [**TABLET-Medium**](https://huggingface.co/datasets/alonsoapp/TABLET-Medium): Includes all examples from _TABLET-Small_, plus **Column Type Annotation**, **Entity Linking**, and **Relation Extraction** tasks. Each task is capped at **140,000 examples**, resulting in a total of **1,117,217 training examples** across **17 tasks**.
+- _(train)_ [**TABLET-Large**](https://huggingface.co/datasets/alonsoapp/TABLET-Large): Includes all examples from _TABLET-Medium_ with **no cap** on task size, resulting in a total of 3,505,311 training examples across 17 tasks.
+- _(dev)_ [**TABLET-dev**](https://huggingface.co/datasets/alonsoapp/TABLET-dev): The **development** set of TABLET.
+- _(test)_ [**TABLET-test**](https://huggingface.co/datasets/alonsoapp/TABLET-test): The **test** set of TABLET.
+
+For more information, see our [paper](https://arxiv.org/pdf/2509.21205), [website](https://precious-panda-5ce815.netlify.app/tablet/), and [GitHub repository](https://github.com/AlonsoApp/TABLET). 
+
+#### Using the Dataset
+
+Given its size, we recommend [streaming](https://huggingface.co/docs/datasets/stream) the dataset instead of downloading it entirely to disk:
+
+```python
+from datasets import load_dataset
+dataset = load_dataset('alonsoapp/TABLET-Small', split='fetaqa_train', streaming=True)
+print(next(iter(dataset)))
+```
+
+#### Data Fields
+
+Each sample within the dataset is structured with the following fields:
+
+*   **`example_id`**: Unique identifier for the example.
+*   **`task`**: The name of the task this example belongs to.
+*   **`src_example_ids`**: IDs of the original examples from the source dataset, formatted as `{"Dataset name": "id"}`. Use the `get_original_example` helper function from [our published code](https://github.com/AlonsoApp/TABLET) to easily retrieve the source example.
+*   **`table_id`**: Unique identifier for the table.
+*   **`table_seed_id`**: ID referencing the table in its original (seed) dataset.
+*   **`table_seed_dataset`**: Name of the dataset where the table originated, typically matching the source dataset of the example.
+*   **`table_page_title`**: For tables sourced from Wikipedia, the corresponding page title.
+*   **`table_section_title`**: For Wikipedia tables, the title of the section where the table appears.
+*   **`table_variant`**: Either "raw" or "highlighted". Some examples visually highlight specific cells and this field indicates whether the table is unmodified (raw) or includes highlights (highlighted).
+*   **`img_source`**: Source of the table image. That is, whether the image comes from the Wikipedia visualization (wikipedia), a synthetic renderization from the data in the soruce dataset (seed_render), or directly copied from the original visualization of the table of the source dataset (PubTabNet, TabMWP). 
+*   **`input`**: The _instructified_ input used for training and evaluation (see [paper](https://arxiv.org/pdf/2509.21205)). The input can be rephrased using information in `metadata`.
+*   **`output`**: The expected model output for the given `input`.
+*   **`split`**: Dataset split: `train`, `dev`, or `test`.
+*   **`metadata`**: Atomic data for the example to enable reconstruction or rephrasing of the instruction. Each key indicates a data element, the value can be obtained from either the _'input'_ or the _'output'_ strings using the substring defined by the character indexes in 'idx'. Use the get_metadata helper function from [our published code](https://github.com/AlonsoApp/TABLET) to retrieve these values.
+*   **`table_wiki_page_id`**: For Wikipedia tables, the page ID corresponding to the article containing the table (useful for Wikipedia API queries).
+*   **`table_wiki_old_id`**: For Wikipedia tables, the “old ID” identifying the article version at the crawl time.
+*   **`table_html`**: HTML representation of the table. Use the `render_table` helper function from [our code](https://github.com/AlonsoApp/TABLET) to render it in its original style. For highlighted variants, highlighted cells use the CSS class `demeter_highlighted_cell`. Remove any decorators for this class in the CSS to render identically to the raw version.
+*   **`table_img`**: The image representation of the table.
+
+#### Citation
+
+If you find **TABLET** useful in your research, please consider citing it by the following BibTeX entry.
+
+```bibtex
+@misc{alonso2025tabletlargescaledatasetrobust,
+      title={TABLET: A Large-Scale Dataset for Robust Visual Table Understanding}, 
+      author={Iñigo Alonso and Imanol Miranda and Eneko Agirre and Mirella Lapata},
+      year={2025},
+      eprint={2509.21205},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2509.21205}, 
+}
+```