---
task_categories:
- feature-extraction
tags:
- table-embeddings
- column-type-annotation
- column-property-annotation
- tabular
---

# TabEmb Representations

Precomputed LLM column embeddings used by the [TabEmb](https://github.com/) pipeline
for column type annotation (CTA) and column property annotation (CPA).

**TODO before publishing:** fill in the sections below, and confirm the licensing
terms of each source benchmark permit redistributing derived embeddings
(SOTAB, T2D, WebTables, WikiTables each have their own license/citation
requirements).

## Contents

Each subdirectory follows `<data_dir_name>/<task>/` (task is `cta` or `cpa`),
containing pickle files named:

```
<task>_<data_name>_<split>_<row_examples>_<model>_<quantized>_<mean_pooling>_<layer_offset>_<layer_pooling>.pkl
```

Datasets included:
- `dataset-sota-dbpedia`
- `dataset-t2d`
- `sotabv2`
- `webtables`
- `wikitable`

## Pickle schema

Each `.pkl` file is a dict with:
- `table_id`: array of table identifiers
- `column_id`: array of column indices
- `embedding`: array of embedding vectors (one per column)
- `label`: array of integer class labels

## Loading

```python
import pickle
with open("dataset-t2d/cta/cta_dataset-t2d_train_25_mistral_True_True_1_none.pkl", "rb") as f:
    data = pickle.load(f)
```

## License / Citation

TODO: state the license for this derived data and cite the original
benchmarks (SOTAB, T2D, WDC WebTables, WikiTables).