ValueError: Object arrays cannot be loaded when allow_pickle=False

The error#

ValueError: Object arrays cannot be loaded when allow_pickle=False

What it means#

You called np.load on a .npy or .npz file that stores Python objects rather than plain numbers. Loading objects requires pickle, Python's object-serialisation format — and unpickling untrusted data can execute arbitrary code. NumPy therefore refuses by default (since version 1.16.3) and makes you opt in.

Why it happens#

Plain numeric arrays save and load without pickle. An array acquires dtype=object when it holds strings, dicts, or ragged rows of different lengths:

np.save("data.npy", np.array([[1, 2], [3, 4, 5]], dtype=object))

Whoever created the file saved objects, knowingly or not. Loading it now needs pickle, so NumPy asks for explicit permission.

How to fix it#

1. If you created the file, or fully trust its source, allow pickle.

data = np.load("data.npy", allow_pickle=True)

That is the entire fix for your own files. Do not do this for files downloaded from the internet or received from strangers — a malicious pickle runs code the moment it loads.

2. For dict-like saves, remember .npz items need .item().

d = np.load("results.npy", allow_pickle=True).item()   # a saved dict comes back

3. Better long-term: stop saving objects. If the array became object by accident — ragged rows are the usual cause — fix the data instead:

arr = np.array(rows)
print(arr.dtype)          # object? your rows have unequal lengths

Pad the rows to equal length, or store each array separately in an .npz:

np.savez("data.npz", images=images, labels=labels)

4. For genuinely structured data, use a format built for it. JSON for nested records, Parquet for tables, safetensors for model weights. All of them load without executing code.

How to prevent it#

Check arr.dtype before saving. If it prints object, understand why before writing the file. Prefer formats that cannot execute code for anything that leaves your machine, and treat allow_pickle=True on a downloaded file as the security decision it is.

Related errors#

• Weights only load failed (torch.load) — PyTorch's version of the same security default
• ValueError: setting an array element with a sequence — how ragged object arrays get created
• git-lfs pointer files instead of weights