load_dataset: TypeError: 'NoneType' object is not callable, on local dataset filename changes

### Describe the bug

I've adapted Databrick's [train_dolly.py](/databrickslabs/dolly/blob/master/train_dolly.py) to train using a local dataset, which has been working. Upon changing the filenames of the `.json` & `.py` files in my local dataset directory, `dataset = load_dataset(path_or_dataset)["train"]` throws the error:

```python
2023-04-30 09:10:52 INFO [training.trainer] Loading dataset from dushowxa-characters
Traceback (most recent call last):
File "/data/dushowxa-dolly/train_dushowxa.py", line 26, in
load_training_dataset()
File "/data/dushowxa-dolly/training/trainer.py", line 89, in load_training_dataset
dataset = load_dataset(path_or_dataset)["train"]
File "/data/dushowxa-dolly/.venv/lib/python3.10/site-packages/datasets/load.py", line 1773, in load_dataset
builder_instance = load_dataset_builder(
File "/data/dushowxa-dolly/.venv/lib/python3.10/site-packages/datasets/load.py", line 1528, in load_dataset_builder
builder_instance: DatasetBuilder = builder_cls(
TypeError: 'NoneType' object is not callable
```
The local dataset filenames were of the form `dushowxa-characters/expanse-dushowxa-characters.json` and are now of the form `dushowxa-characters/dushowxa-characters.json` (the word `expanse-` was removed from the filenames). Is this perhaps a dataset caching issue?

I have attempted to manually clear caches, but to no effect:

```sh
rm -rfv ~/.cache/huggingface/datasets/*
rm -rfv ~/.cache/huggingface/modules/*
```

### Steps to reproduce the bug

Run `python3 train_dushowxa.py` (adapted from Databrick's [train_dolly.py](/databrickslabs/dolly/blob/master/train_dolly.py)).

### Expected behavior

Training succeeds as before local dataset filenames were changed.

### Environment info

Ubuntu 22.04, Python 3.10.6, venv
```python
accelerate>=0.16.0,<1
click>=8.0.4,<9
datasets>=2.10.0,<3
deepspeed>=0.9.0,<1
transformers[torch]>=4.28.1,<5
langchain>=0.0.139
```