metadata

datasets:
  - PKU-Alignment/PKU-SafeRLHF
language:
  - en
tags:
  - reinforcement-learning-from-human-feedback
  - reinforcement-learning
  - beaver
  - safety
  - llama
  - ai-safety
  - deepspeed
  - rlhf
  - alpaca
library_name: safe-rlhf

🦫 Beaver's Cost Model

Model Details

The Beaver cost model is a preference model trained using the PKU-SafeRLHF dataset. It can play a role in the safe RLHF algorithm, helping the Beaver model become more safe and harmless.

Developed by: the PKU-Alignment Team.
Model Type: An auto-regressive language model based on the transformer architecture.
License: Non-commercial license.
Fine-tuned from model: LLaMA, Alpaca.

Model Sources

Repository: https://github.com/PKU-Alignment/safe-rlhf
Beaver: https://huggingface.co/PKU-Alignment/beaver-7b-v3.0
Dataset: https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF
Reward Model: https://huggingface.co/PKU-Alignment/beaver-7b-unified-reward
Cost Model: https://huggingface.co/PKU-Alignment/beaver-7b-unified-cost
Dataset Paper: https://arxiv.org/abs/2307.04657
Paper: https://arxiv.org/abs/2310.12773

How to Use the Cost Model

import torch
from transformers import AutoTokenizer
from safe_rlhf.models import AutoModelForScore

model = AutoModelForScore.from_pretrained('PKU-Alignment/beaver-7b-unified-cost', torch_dtype=torch.bfloat16, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained('PKU-Alignment/beaver-7b-unified-cost')

input = 'BEGINNING OF CONVERSATION: USER: hello ASSISTANT:Hello! How can I help you today?'

input_ids = tokenizer(input, return_tensors='pt')
output = model(**input_ids)
print(output)

# ScoreModelOutput(
#     scores=tensor([[[-2.7656],
#          [ 0.8320],
#          [-2.7656],
#          [-2.7500],
#          [-0.9023],
#          [-0.7891],
#          [-0.3125],
#          [-0.8008],
#          [-0.5117],
#          [-1.1562],
#          [-2.3906],
#          [-1.2266],
#          [-1.1797],
#          [-3.3281],
#          [-4.4062],
#          [-1.0234],
#          [-1.1484],
#          [-2.1406],
#          [-2.9531],
#          [-4.6250],
#          [-4.5312],
#          [-3.3594],
#          [-4.1250],
#          [-3.0156],
#          [-3.5156],
#          [-5.0000],
#          [-5.7812],
#          [-7.6562]]], grad_fn=<ToCopyBackward0>),
#     end_scores=tensor([[-7.6562]], grad_fn=<ToCopyBackward0>),
#     last_hidden_state=tensor([[[ 0.7148,  0.3594, -1.0234,  ...,  0.5039, -0.0737,  1.4375],
#          [ 1.0781, -1.2812,  1.5078,  ...,  0.9102,  1.3594,  1.4141],
#          [ 0.8047,  0.4551, -0.3262,  ...,  0.3887,  0.6484, -0.4629],
#          ...,
#          [-0.1836, -0.6094, -0.8086,  ..., -0.5078,  0.8086,  1.1719],
#          [ 0.9727, -1.5156, -1.2656,  ..., -0.9766,  0.3535,  1.0156],
#          [ 4.2812, -1.6797, -0.4238,  ...,  0.6758, -1.1875, -1.1562]]],
#        dtype=torch.bfloat16, grad_fn=<ToCopyBackward0>),
#     end_last_hidden_state=tensor([[ 4.2812, -1.6797, -0.4238,  ...,  0.6758, -1.1875, -1.1562]],
#        dtype=torch.bfloat16, grad_fn=<ToCopyBackward0>),
#     end_index=tensor([27])
# )