TabGenBench / README.md
aodongl's picture
Add LICENSE
6ad3fe0 verified
|
Raw History Blame Contribute Delete
2.36 kB
---
pretty_name: TabGenBench downstream task masks
tags:
- tabular
- benchmark
- imputation
task_categories:
- tabular-classification
- tabular-regression
size_categories:
- 1M<n<10M
license: apache-2.0
---
# TabGenBench downstream task masks
This repository stores the fixed masks of the [TabGenBench](https://github.com/granica-ai/tabgen-bench/) downstream tasks:
for each of the benchmark's 28 tables, which features of the test split are masked
for prediction or imputation tasks.
This repository holds **masks only**. A mask is a list of `[row, column]` positions
of the prepared test split in TabGenBench. The package downloads and checks them on
first use automatically:
```python
import tabgen_bench as tgb
tgb.list_downstream_tasks("sdss_17")
# ('prediction', 'imputation_30', 'imputation_50')
task = tgb.load_downstream_task("sdss_17", "imputation_50")
observed, targets = task["observed"], task["targets"]
```
`observed` sets the hidden cells to `np.nan`; `targets` reveals those values.
After download, the folder structure looks like
```
<data root>/
├── downstream/ from this repository
├── prepared/ built by the package
├── raw/ the original table, from TabArena/BeyondArena
└── table_sources/ the package's verification records
```
## Tasks
Every table has the same three tasks.
| Task | What is hidden |
| --- | --- |
| `prediction` | The associated supervised targeting column |
| `imputation_30` | 30% columns, chosen at random for each row |
| `imputation_50` | 50% columns, chosen at random for each row |
## Files
```
downstream/<table>.npz the positions, one array per task
downstream/<table>.json the manifest that describes and binds them
```
There are 56 files, 83 MB in all.
`<table>.npz`:
One array per task, named `positions_0`, `positions_1`, `positions_2`,
representing `prediction`, `imputation_30`, `imputation_50`.
Each array has shape `(hidden cells, 2)`. A row of it is
`[row_index, column_index]` for the prepared test split.
The arrays are plain numeric data and load with `allow_pickle=False`.
`<table>.json`: metadata and checksums for the associated masks.
## Licence
The tables themselves stay under the licences of their original sources and
are not redistributed here.