How to create numpy records array with numerical entries without dtype name

There are several items in your question that require explanation.

1. The (3,) makes the object an array of length 3. That's why you see '|V3'.
2. The "garbage values" you see are the random values used to initialize the array when was created (because you created an empty array). You won't see this behavior if you initialize with values (or zeros or ones).
3. When reading HDF5 data you don't have to create the NumPy array in advance. You can create it when you read the data.
4. Better still, you can create a h5py dataset object to access the data, then use it "like" an array. Dataset objects behave like NumPy objects (and use less memory). For example, you can get the dtype and shape. These will be the same as the equivalent NumPy array.

Example to access values in an HDF5 dataset as a dataset object or an array:

with h5py.File('example.h5') as h5f:
    # create dataset object:
    ds_obj = h5f['dataset_name']
    print(ds_obj.dtype)
    print(ds_obj.shape)
    # create an array of the entire dataset:
    arr = h5f['dataset_name'][()]
    # or:
    arr = ds_obj[()]
    # create an array from a slice of the dataset:
    arr_slice = h5f['dataset_name'][0:10,]
    # or:
    arr_slice = ds_obj[0:10]

Example to create the record array (for completeness):

rec_arr = np.empty((10,), dtype=([('int_value','uint8'), ('str_value','S10')]))
rec_arr[0,] = (100, 'row 0 val')