sileod

/

deberta-v3-large-tasksource-rlhf-reward-model

like 11

I can't reproduce it.

this is my test notebook

import torch

torch.set_grad_enabled(False)
device = "cuda:0"

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "sileod/deberta-v3-large-tasksource-rlhf-reward-model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval().to(device)

from datasets import load_dataset

dataset = load_dataset("heegyu/hh-rlhf-vicuna-format", split="test")

def join_conv(convs):
    lines = []
    for conv in convs:
        if conv["from"] == "human":
            lines.append("Human: " + conv["value"])
        else:
            lines.append("Assistant: " + conv["value"])
    return "\n\n".join(lines)

def rank(context, response):
    inputs = tokenizer(context, response, truncation=True, return_tensors='pt').to(device)
    return model(**inputs).logits[0].cpu().detach().tolist()[0]

def map_item(item):
    # with context
    # context = join_conv(item["context"]) + "\n\n" + join_conv([item["instruction"]])
    # context = context.strip()
    
    # without context
    context = item["instruction"]["value"]


    chosen = rank(context, item["chosen"]["value"])
    rejected = rank(context, item["rejected"]["value"])
    
    return {
        "chosen_score": chosen,
        "rejected_score": rejected,
        "predict": 1 if chosen > rejected else 0
    }


tokenizer.truncation_side = "left"

classified = dataset.map(map_item, load_from_cache_file=False)
print(classified.to_pandas().mean()["predict"])

and result is

• with context: 0.6792563143124415
• without context: 0.6627689429373246