safe-rlhf

Beaver is a highly modular open-source RLHF framework developed by the PKU-Alignment team at Peking University. It aims to provide training data and a reproducible code pipeline for alignment research, especially constrained alignment LLM research via Safe RLHF methods.

The key features of Beaver are:

• Support SFT, RLHF and Safe RLHF training for popular pre-trained models: LLaMA, OPT, Baichuan, etc.
• Provide a large human-labeled dataset (up to 1M pairs) including both helpful and harmless preferences to support reproducible RLHF research.
• Support training for Reward Model & Cost Model, and provide pre-trained checkpoints.
• Support customized parameters and datasets for SFT and RLHF.
• Provide multi-scale metrics for safety constraints verification, e.g., BIG-bench, GPT-4 Evaluation.

🦫 What's New?

• 🎉 2024/06/13: We are pleased to announce the open-sourcing of our PKU-SafeRLHF dataset version 1.0. This release advances over the initial beta version by incorporating human-AI joint annotations, expanding the scope of harm categories, and introducing detailed severity level labels. For further details and access, please visit our dataset page on 🤗 Hugging Face: PKU-Alignment/PKU-SafeRLHF.
• 🎉 2024/01/16: Our method Safe RLHF has been accepted by ICLR 2024 Spotlight.
• 📄 2023/10/19: We've released our Safe RLHF paper on arXiv, detailing our new safe alignment algorithm and its implementation.
• 🚀 2023/07/10: We're delighted to announce the open-sourcing of Beaver-7B v1 / v2 / v3 models as the first milestone of the Safe RLHF training series, complemented by the corresponding Reward Models v1 / v2 / v3 / unified and Cost Models v1 / v2 / v3 / unified checkpoints on 🤗 Hugging Face.
• 🔥 2023/07/10: We extend the open-source safety preference dataset, PKU-Alignment/PKU-SafeRLHF, which now contains over 300k examples. (See also section PKU-SafeRLHF-Dataset)
• ⚙ 2023/07/05: We enhanced our support for Chinese pre-training models and incorporated additional open-source Chinese datasets. (See also sections Chinese Support (中文支持) and Custom Datasets (自定义数据集))
• ⭐️ 2023/05/15: First release of the Safe RLHF pipeline, evaluation results, and training code.

Table of Contents

• Constrained Value Alignment via Safe RLHF
• Comparison with Other RLHF Libraries
• PKU-SafeRLHF-Dataset
  • PKU-SafeRLHF-10K
  • PKU-SafeRLHF-1M
• Why "Beaver"
• Beaver vs. Alpaca
• Installation
• Training
• Custom Datasets
• Inference
  • Interactive CLI Demo
  • Interactive Arena
• Chinese Support (中文支持)
• Benchmark and Evaluation
  • Arena via Reward and Cost Models
  • BIG-bench
  • GPT-4 Evaluation
• Future Plans
• Citation
• PKU-Alignment Team
• License

Constrained Value Alignment via Safe RLHF

Reinforcement Learning from Human Feedback: reward maximization via preference learning

Safe Reinforcement Learning from Human Feedback: constrained reward maximization via preference learning

where $R (\cdot)$ and $C (\cdot)$ are reward and cost functions respectively. They are neural networks known as human proxies trained on human preferences.

The ultimate goal is to find a model $\pi_{\theta}$ that is both helpful (high reward) and harmless (low cost).

Comparison with Other RLHF Libraries

Compare with other frameworks supporting RLHF, safe-rlhf is the first framework to support all stages from SFT to RLHF and Evaluation. In addition, safe-rlhf is the first framework that takes safety preference under consideration during the RLHF stage. It holds a more theoretical guarantee for constrained parameter searching in the policy space.

	SFT	Preference Model1 Training	RLHF	Safe RLHF	PTX Loss	Evaluation	Backend
Beaver(Safe-RLHF)	✔️	✔️	✔️	✔️	✔️	✔️	DeepSpeed
trlX	✔️	❌2	✔️	❌	❌	❌	Accelerate / NeMo
DeepSpeed-Chat	✔️	✔️	✔️	❌	✔️	❌	DeepSpeed
Colossal-AI	✔️	✔️	✔️	❌	✔️	❌	ColossalAI
AlpacaFarm	❌3	✔️	✔️	❌	❌	✔️	Accelerate

PKU-SafeRLHF-Dataset

The PKU-SafeRLHF dataset is a human-labeled dataset containing both performance and safety preferences. It includes constraints in over ten dimensions, such as insults, immorality, crime, emotional harm, and privacy, among others. These constraints are designed for fine-grained value alignment in RLHF technology.

To facilitate multi-round fine-tuning, we will release the initial parameter weights, required datasets, and training parameters for each round. This ensures reproducibility in scientific and academic research. The dataset will be released gradually through rolling updates.

The dataset is available on Hugging Face: PKU-Alignment/PKU-SafeRLHF.

PKU-SafeRLHF-10K

PKU-SafeRLHF-10K is a subset of PKU-SafeRLHF that contains the first round of Safe RLHF training data with 10K instances, including safety preferences. You can find it on Hugging Face: PKU-Alignment/PKU-SafeRLHF-10K.

PKU-SafeRLHF-1M

We will gradually release the full Safe-RLHF datasets, which include 1M human-labeled pairs for both helpful and harmless preferences.

Why "Beaver"

Beaver is a large language model based on LLaMA, trained using safe-rlhf. It is developed upon the foundation of the Alpaca model, by collecting human preference data related to helpfulness and harmlessness and employing the Safe RLHF technique for training. While maintaining the helpful performance of Alpaca, Beaver significantly improves its harmlessness.

Beavers are known as the "natural dam engineers" as they are adept at using branches, shrubs, rocks, and soil to build dams and small wooden houses, creating wetland environments suitable for other creatures to inhabit, making them an indispensable part of the ecosystem. To ensure the safety and reliability of Large Language Models (LLMs) while accommodating a wide range of values across different populations, the Peking University team has named their open-source model "Beaver" and aims to build a dam for LLMs through the Constrained Value Alignment (CVA) technology. This technology enables fine-grained labeling of information and, combined with secure reinforcement learning methods, significantly reduces model bias and discrimination, thereby enhancing the model's safety. Analogous to the role of beavers in the ecosystem, the Beaver model will provide crucial support for the development of large language models and make positive contributions to the sustainable development of artificial intelligence technology.

Beaver vs. Alpaca

Following the evaluation methodology of the Vicuna model, we utilized GPT-4 to evaluate Beaver. The results indicate that, compared to Alpaca, Beaver exhibits significant improvements in multiple dimensions related to safety.

Significant distribution shift for safety preferences after utilizing the Safe RLHF pipeline on the Alpaca-7B model.

Installation

Clone the source code from GitHub:

git clone https://github.com/PKU-Alignment/safe-rlhf.git
cd safe-rlhf

Native Runner: Setup a conda environment using conda / mamba:

conda env create --file conda-recipe.yaml  # or `mamba env create --file conda-recipe.yaml`

This will automatically setup all dependencies.

Containerized Runner: Other than using the native machine with conda isolation, as an alternative, you can also use docker images to configure the environment.

Firstly, please follow NVIDIA Container Toolkit: Installation Guide and NVIDIA Docker: Installation Guide to setup nvidia-docker. Then you can run:

make docker-run

This command will build and start a docker container installed with proper dependencies. The host path / will be mapped to /host and the current working directory will be mapped to /workspace inside the container.

Training

safe-rlhf supports a complete pipeline from Supervised Fine-Tuning (SFT) to preference model training to RLHF alignment training.

0. Follow the instructions in section Installation to setup the training environment properly.

conda activate safe-rlhf
export WANDB_API_KEY="..."  # your W&B API key here

or

make docker-run
export WANDB_API_KEY="..."  # your W&B API key here

1. Supervised Fine-Tuning (SFT)

bash scripts/sft.sh \
    --model_name_or_path <your-model-name-or-checkpoint-path> \
    --output_dir output/sft

NOTE: You may need to update some of the parameters in the script according to your machine setup, such as the number of GPUs for training, the training batch size, etc.

2. Value Models (reward model & cost model)

bash scripts/reward-model.sh \
    --model_name_or_path output/sft \
    --output_dir output/rm

bash scripts/cost-model.sh \
    --model_name_or_path output/sft \
    --output_dir output/cm

3. RLHF (Optional)

bash scripts/ppo.sh \
    --actor_model_name_or_path output/sft \
    --reward_model_name_or_path output/rm \
    --output_dir output/ppo

4. Safe-RLHF

bash scripts/ppo-lag.sh \
    --actor_model_name_or_path output/sft \
    --reward_model_name_or_path output/rm \
    --cost_model_name_or_path output/cm \
    --output_dir output/ppo-lag

An example of commands to run the whole pipeline with LLaMA-7B:

conda activate safe-rlhf
bash scripts/sft.sh --model_name_or_path ~/models/llama-7b --output_dir output/sft
bash scripts/reward-model.sh --model_name_or_path output/sft --output_dir output/rm
bash scripts/cost-model.sh --model_name_or_path output/sft --output_dir output/cm
bash scripts/ppo-lag.sh \
    --actor_model_name_or_path output/sft \
    --reward_model_name_or_path output/rm \
    --cost_model_name_or_path output/cm \
    --output_dir output/ppo-lag

Computational Requirements

All training processes listed above are tested with LLaMA-7B on a cloud server with 8 x NVIDIA A800-80GB GPUs.

Users, who do not have enough GPU memory resources, can enable DeepSpeed ZeRO-Offload to alleviate the peak GPU memory usage.

All training scripts can pass with an extra option --offload (defaults to none, i.e., disable ZeRO-Offload) to offload the tensors (parameters and/or optimizer states) to CPU. For example:

bash scripts/sft.sh \
    --model_name_or_path ~/models/llama-7b \
    --output_dir output/sft \
    --offload all  # or `parameter` or `optimizer`

For multi-node settings, users can refer to the DeepSpeed: Resource Configuration (multi-node) documentation for more details. Here is an example to start the training process on 4 nodes (each has 8 GPUs):

# myhostfile
worker-1 slots=8
worker-2 slots=8
worker-3 slots=8
worker-4 slots=8

Then launch the training scripts with:

bash scripts/sft.sh \
    --hostfile myhostfile \
    --model_name_or_path ~/models/llama-7b \
    --output_dir output/sft

Custom Datasets

safe-rlhf provides an abstraction to create datasets for all of the Supervised Fine-Tuning, preference model training, and RL training stages.

class RawSample(TypedDict, total=False):
    """Raw sample type.

    For SupervisedDataset, should provide (input, answer) or (dialogue).
    For PreferenceDataset, should provide (input, answer, other_answer, better).
    For SafetyPreferenceDataset, should provide (input, answer, other_answer, safer, is_safe, is_other_safe).
    For PromptOnlyDataset, should provide (input).
    """

    # Texts
    input: NotRequired[str]  # either `input` or `dialogue` should be provided
    """User input text."""
    answer: NotRequired[str]
    """Assistant answer text."""
    other_answer: NotRequired[str]
    """Other assistant answer text via resampling."""
    dialogue: NotRequired[list[str]]  # either `input` or `dialogue` should be provided
    """Dialogue history."""

    # Flags
    better: NotRequired[bool]
    """Whether ``answer`` is better than ``other_answer``."""
    safer: NotRequired[bool]
    """Whether ``answer`` is safer than ``other_answer``."""
    is_safe: NotRequired[bool]
    """Whether ``answer`` is safe."""
    is_other_safe: NotRequired[bool]
    """Whether ``other_answer`` is safe."""

Here is an example to implement a custom dataset (see safe_rlhf/datasets/raw for more examples):

import argparse
from datasets import load_dataset
from safe_rlhf.datasets import RawDataset, RawSample, parse_dataset


class MyRawDataset(RawDataset):
    NAME = 'my-dataset-name'

    def __init__(self, path=None) -> None:
        # Load a dataset from Hugging Face
        self.data = load_dataset(path or 'my-organization/my-dataset')['train']

    def __getitem__(self, index: int) -> RawSample:
        data = self.data[index]
        # Construct a `RawSample` dictionary from your custom dataset item
        return RawSample(
            input=data['col1'],
            answer=data['col2'],
            other_answer=data['col3'],
            better=float(data['col4']) > float(data['col5']),
            ...
        )

    def __len__(self) -> int:
        return len(self.data)  # dataset size


def parse_arguments():
    parser = argparse.ArgumentParser(...)
    parser.add_argument(
        '--datasets',
        type=parse_dataset,
        nargs='+',
        metavar='DATASET[:PROPORTION[:PATH]]',
    )
    ...
    return parser.parse_args()


def main():
    args = parse_arguments()
    ...


if __name__ == '__main__':
    main()

Then you can pass this dataset to the training scripts as:

python3 train.py --datasets my-dataset-name

You may also pass multiple datasets with optionally additional dataset proportions (separated by a colon :). For example:

python3 train.py --datasets alpaca:0.75 my-dataset-name:0.5

This will use randomly split 75% of the Stanford Alpaca dataset and 50% of your custom dataset.

In addition, the dataset argument can also be followed by a local path (separated by a colon :) if you have already cloned the dataset repository from Hugging Face.

git lfs install
git clone https://huggingface.co/datasets/my-organization/my-dataset ~/path/to/my-dataset/repository
python3 train.py --datasets alpaca:0.75 my-dataset-name:0.5:~/path/to/my-dataset/repository

NOTE: The dataset class must be imported before the training script begins to parse the command line arguments.

Inference

Interactive CLI Demo

python3 -m safe_rlhf.serve.cli --model_name_or_path output/sft  # or output/ppo-lag

Interactive Arena

python3 -m safe_rlhf.serve.arena --red_corner_model_name_or_path output/sft --blue_corner_model_name_or_path output/ppo-lag

Chinese Support (中文支持)

The Safe-RLHF pipeline supports not only the LLaMA model family but also other pre-trained models such as Baichuan, InternLM, etc. that offer better support for Chinese. You just need to update the path to the pre-trained model in the training and inference code.

Safe-RLHF 管道不仅仅支持 LLaMA 系列模型，它也支持其他一些对中文支持更好的预训练模型，例如 Baichuan 和 InternLM 等。你只需要在训练和推理的代码中更新预训练模型的路径即可。

# SFT training
bash scripts/sft.sh --model_name_or_path baichuan-inc/Baichuan-7B --output_dir output/baichuan-sft

# Inference
python3 -m safe_rlhf.serve.cli --model_name_or_path output/baichuan-sft

In the meantime, we've added support for Chinese datasets such as the Firefly and MOSS series to our raw-datasets. You only need to change the dataset path in the training code to use the corresponding dataset for fine-tuning the Chinese pre-training model:

同时，我们也在 raw-datasets 中增加了支持一些中文数据集，例如 Firefly 和 MOSS 系列等。在训练代码中更改数据集路径，你就可以使用相应的数据集来微调中文预训练模型：

# scripts/sft.sh
-	--train_datasets alpaca \
+	--train_datasets firefly \

For instructions on how to add custom datasets, please refer to section Custom Datasets.

关于如何添加自定义数据集的方法，请参阅章节 Custom Datasets (自定义数据集)。

Benchmark and Evaluation

Arena via Reward and Cost Models

scripts/arena-evaluation.sh \
    --red_corner_model_name_or_path output/sft \
    --blue_corner_model_name_or_path output/ppo-lag \
    --reward_model_name_or_path output/rm \
    --cost_model_name_or_path output/cm \
    --output_dir output/arena-evaluation

BIG-bench

# Install BIG-bench
git clone https://github.com/google/BIG-bench.git
(
    cd BIG-bench
    python3 setup.py sdist
    python3 -m pip install -e .
)

# BIG-bench evaluation
python3 -m safe_rlhf.evaluate.bigbench \
    --model_name_or_path output/ppo-lag \
    --task_name <BIG-bench-task-name>

GPT-4 Evaluation

# Install OpenAI Python API
pip3 install openai
export OPENAI_API_KEY="..."  # your OpenAI API key here

# GPT-4 evaluation
python3 -m safe_rlhf.evaluate.gpt4 \
    --red_corner_model_name_or_path output/sft \
    --blue_corner_model_name_or_path output/ppo-lag

Future Plans

• Beaver-7B checkpoint is released on Hugging Face.
• Release Safe RLHF paper preprint.
• We will gradually release the full Safe-RLHF datasets.
• Train Larger LLM with Safe-RLHF.
• Support memory-efficient training, such as LoRA, PEFT, etc.

Citation

If you find Safe-RLHF useful or use Safe-RLHF (model, code, dataset, etc.) in your research, please consider citing the following work in your publications.

@inproceedings{safe-rlhf,
  title={Safe RLHF: Safe Reinforcement Learning from Human Feedback},
  author={Josef Dai and Xuehai Pan and Ruiyang Sun and Jiaming Ji and Xinbo Xu and Mickel Liu and Yizhou Wang and Yaodong Yang},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=TyFrPOKYXw}
}
@inproceedings{beavertails,
  title={BeaverTails: Towards Improved Safety Alignment of {LLM} via a Human-Preference Dataset},
  author={Jiaming Ji and Mickel Liu and Juntao Dai and Xuehai Pan and Chi Zhang and Ce Bian and Boyuan Chen and Ruiyang Sun and Yizhou Wang and Yaodong Yang},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2023},
  url={https://openreview.net/forum?id=g0QovXbFw3}
}

PKU-Alignment Team

All students below contributed equally and the order is determined alphabetically:

• Juntao Dai
• Jiaming Ji
• Xuehai Pan
• Ruiyang Sun

All advised by Yizhou Wang and Yaodong Yang. Acknowledge: We appreciate Ms. Yi Qu for designing the Beaver logo.

Acknowledgment

This repository benefits from LLaMA, Stanford Alpaca, DeepSpeed, and DeepSpeed-Chat. Thanks for their wonderful works and their efforts for democratizing the LLM research. Safe-RLHF and its related assets are built and open-sourced with love 🤗❤️.

This work is supported and funded by the Peking University.

License

Safe-RLHF is released under Apache License 2.0.