Model for testing RM scripts

This model is just GPT2 base (~100M param) with a value head appended, untrained. Use this for debugging RLHF setups (could make a smaller one too). The predictions should be somewhat random.

Load the model as follows:

from transformers import AutoModelForSequenceClassification
rm = AutoModelForSequenceClassification.from_pretrained("natolambert/gpt2-dummy-rm")

or as a pipeline

from Transformers import pipeline
reward_pipe = pipeline(
        "text-classification",
        model="natolambert/gpt2-dummy-rm",
        # revision=args.model_revision,
        # model_kwargs={"load_in_8bit": True, "device_map": {"": current_device}, "torch_dtype": torch.float16},
    )
reward_pipeline_kwargs = {}
pipe_outputs = reward_pipe(texts, **reward_pipeline_kwargs)