`PandasArrayExtensionDtype._metadata` should be a tuple, not a string

[kohankhaki]

Describe the bug

There's a small typo in PandasArrayExtensionDtype (src/datasets/features/features.py):

_metadata = "value_type"

Pandas expects _metadata to be a tuple of attribute names, like ("value_type",), not a plain string (API documentation ref). Because it's a string, pandas iterates over each character in "value_type" when comparing dtypes. So it tries to read attributes named v, a, l, u, e, etc, and crashes on the first one:

AttributeError: 'PandasArrayExtensionDtype' object has no attribute 'v'

This shows up whenever:

1. Write a parquet file with array columns (e.g. array[float64] from Array2D/Array3D features)
2. Read it back with pd.read_parquet() after import datasets
3. Filter rows with a boolean mask like df[mask]

I ran into this while splitting a LeRobot dataset, but the bug itself is in datasets. I think the same root cause is discussed in an issue in LeRobot repo: huggingface/lerobot#2445.

Steps to reproduce the bug

You can reproduce the error using the code below:

import tempfile
from pathlib import Path

import datasets
import pandas as pd
from datasets import Features, Array2D

path = Path(tempfile.mkdtemp()) / "test.parquet"

data = {
    "episode_index": [0, 0, 1, 1],
    "values": [
        [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]],
        [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
        [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]],
        [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
    ],
}
features = Features({
    "episode_index": datasets.Value("int64"),
    "values": Array2D(shape=(2, 3), dtype="float64"),
})

ds = datasets.Dataset.from_dict(data, features=features)
ds.to_parquet(path)

df2 = pd.read_parquet(path)
print(df2["values"].dtype)

mask = df2["episode_index"].isin([0])
df2[mask] 

output:

AttributeError: 'PandasArrayExtensionDtype' object has no attribute 'v'

Expected behavior

df[mask] should work on DataFrames with array[float64] columns read from parquet

Environment info

• datasets version: 5.0.2.dev0
• Platform: Linux-6.8.0-136-generic-x86_64-with-glibc2.39
• Python version: 3.12.13
• huggingface_hub version: 1.26.0
• PyArrow version: 25.0.0
• Pandas version: 3.0.5
• fsspec version: 2026.6.0