Local dataset loading in reward model server

[aws-jiadingg]

In the RL training README.md, it mentions that we can specify dataset_path to load from a json dataset, i.e.:

bash scripts/train_reinforce_ray_rule_rm.sh # reinforcement++
...
# dataset_path="./data/acecode_87K/acecode_87K.json"
...

However, in the implementation, the dataset loading statement appears to only support loading from huggingface hub (missing the path="json" argument necessary to load from local?)

https://github.com/jdf-prog/OpenRLHF/blob/b2e2cdb4bcaa68baa73ed46e025e1051feed5e20/openrlhf/cli/serve_rm.py#L165

To properly load from a local JSON file, should the code be something like?:

# For local JSON files
if os.path.exists(args.dataset):
    self.dataset = datasets.load_dataset("json", data_files=args.dataset, split="train")
else:
    # Fallback to Hugging Face datasets
    self.dataset = datasets.load_dataset(args.dataset, split="train")