This repository contains the code and released models for our paper SimPO: Simple Preference Optimization with a Reference-Free Reward. We propose a simpler and more effective preference optimization algorithm than DPO (Direct Preference Optimization) without using a reference model. SimPO outperforms DPO and its latest variants across AlpacaEval 2, MT-Bench, and Arena-Hard benchmarks under various settings.

• [2024.10.12] To facilitate reproducibility, we release the training curves for Llama3-Instruct and Gemma2-IT:
  • Llama3-Instruct-SimPO
  • Llama3-Instruct-SimPO v0.2
  • Gemma2-IT-SimPO
• [2024.07.17] We released a new SimPO model gemma-2-9b-it-SimPO by fine-tuning Google's gemma-2 9B model using on-policy UltraFeedback data annotated by ArmoRM, achieving a 72.4 LC win rate on AlpacaEval 2 (#1 on the Leaderboard 🎉🎉) and a 59.1 win rate on Arena-Hard! Please find the training script here and the data generation scripts here!
• [2024.07.08] We updated our paper (v2)
  • Additional baselines (RRHF, SLiC-HF, CPO)
  • New Llama3-Instruct setting (v0.2) with ArmoRM as the preference label annotator, yielding a better-performing model, Llama-3-Instruct-8B-SimPO-v0.2, with a 53.7 LC win rate on AlpacaEval 2 and a 36.5 win rate on Arena-Hard (training script)!
  • SimPO trainer update for better reproducibility. The hyperparameter gamma changed to gamma_beta_ratio for easier tuning.

• SimPO: Simple Preference Optimization with a Reference-Free Reward
  • Changelog
  • Tips for Running SimPO
  • Released Models
  • Install Requirements
  • Training scripts
  • Evaluation
  • Bugs or Questions?
  • Citation

Given the various inquiries about SimPO, we provide a list of tips to help you reproduce our paper results and achieve better outcomes for running SimPO on your own tasks.

We provide an environment file including the python package versions we used in our experiments. For optimal reproducibility, we recommend using the same package versions. However, please note that results may still vary due to differences in hardware configurations and CUDA versions, etc.

Hyperparameter tuning is crucial for SimPO (and other preference optimization algorithms in general). The three main hyperparameters of SimPO to focus on are learning_rate, beta, and gamma (we recommend keeping the total batch size fixed at 128).

• learning_rate: It is the most critical hyperparameter for preference optimization. A large learning rate (e.g., 1e-5) can significantly degrade performance, causing the model to produce incoherent sentences or completely repetitive responses. We recommend grid searching over 3e-7, 5e-7, 8e-7, and 1e-6, if resources allow. We find that a smaller learning rate (e.g., 5e-7) is more suitable for reasoning intensive domains like math for both DPO and SimPO.
• beta: Beta controls the reward scaling between winning and losing responses. SimPO requires a much larger beta than DPO. In our preprint, we used a beta of 2.0 or 2.5, but in many cases, an even larger beta (e.g., 10) could yield better results.
• gamma: Gamma controls the target reward margin. We suggest tuning the ratio of gamma to beta (i.e., gamma / beta). We recommend using 0.5 as a starting point for gamma_beta_ratio and grid searching between 0 and 1. A well-tuned gamma_beta_ratio can provide a modest improvement, but it is not as critical as other hyperparameters.

We used the following hyperparameters for training the released models (note that in our latest update, we changed the hyperparameter gamma to gamma_beta_ratio as the latter is normalized and easier to tune under different beta values).

For DPO, the best hyperparameters for each setting are as follows.

Our released Llama3 models use the initial version of the Llama3 tokenizer (prior to this PR). We have found that the updated Llama3 tokenizer with vLLM occasionally introduces two BOS tokens, which can affect evaluation results. Therefore, please ensure that only one BOS token is included in the prompt after applying the Llama3 chat template during any evaluation.

Notably, if you are training Llama3 and evaluating the trained models on AlpacaEval 2 and Arena-Hard using the templates provided in this repo, please make sure to use the pre-update Llama3 tokenizer (i.e., the one before the PR).

Please make sure that you use alpaca-eval==0.6.2 and model configurations in our repo for successfully reproducing AlpacaEval 2 results. AlpacaEval has a major revision for vllm decoding since 0.6.3 and causes a discrepancy from our experiments.

The CPO_SIMPO repository did preliminary experiments and observed that in some cases, adding an additional SFT loss can help improve results. In our own experiments, the SFT regularization helps preserve the reasoning ability (e.g., GSM8K) but degrades chat performance. If you'd like to apply SFT regularization, you can set sft_weight to be a positive value (by default it's 0).

We release the following two models that are built on top of the strong google/gemma-2-9b-it model by training DPO and SimPO on the on-policy dataset princeton-nlp/gemma2-ultrafeedback-armorm. For GSM and MMLU, we use the ZeroEval repository which aims to evaluate instruction-tuned LLMs (i.e., chat models instead of base models) for their zero-shot performance on reasoning and knowledge heavy tasks. More results on WildBench are coming soon.

• Compared to the llama3 models, we found that the gemma models exhibit significantly less catastrophic forgetting on math tasks (e.g., GSM) and MMLU, despite the ultrafeedback dataset having limited math-related data. This demonstrates that the google/gemma-2-9b-it model is more suitable for continued preference optimization.
• SimPO and DPO perform comparably across all benchmarks, but SimPO is inherently simpler and less resource-intensive.

We found that using a strong reward model for annotating preference optimization datasets is crucial. In this iteration, we have reannotated the dataset princeton-nlp/llama3-ultrafeedback-armorm using a more powerful reward model, RLHFlow/ArmoRM-Llama3-8B-v0.1. As a result, the v0.2 models demonstrate significantly improved performance compared to the v0.1 models.

Caveat: We have observed that the SimPO v0.2 model often struggles with generating outputs that require adherence to specific structures, such as json. This issue arises from a combination of factors: the llama3-instruct model's tendency to forget and the large learning rate (e.g., 1e-6) used during training, which causes deviation from the original model. To address this, we developed SimPO models based on the google/gemma-2-9b-it. We found that changing the initial model significantly mitigates the forgetting issue and reduces the impact of the learning rate.

Below is the complete list of models evaluated in our preprint. We used the HuggingFaceH4/ultrafeedback_binarized dataset to train the Mistral Base and Llama3 Base models, the princeton-nlp/mistral-instruct-ultrafeedback dataset to train the Mistral Instruct models, and the princeton-nlp/llama3-ultrafeedback dataset to train the Llama3 Instruct models. The latter two datasets are annotated by the llm-blender/PairRM model.

Please refer to the generate.py script for detailed instructions on loading the model with the appropriate chat template.

Our codebase is built upon the alignment-handbook repo. The following steps will guide you through the installation process.

First, create a Python virtual environment using e.g. Conda:

conda create -n handbook python=3.10 && conda activate handbook

Next, install PyTorch v2.2.2. Since this is hardware-dependent, we direct you to the PyTorch Installation Page.

You can then install the remaining package dependencies of alignment-handbook as follows:

git clone https://github.com/huggingface/alignment-handbook.git
cd ./alignment-handbook/
python -m pip install .

You will also need Flash Attention 2 installed, which can be done by running:

python -m pip install flash-attn --no-build-isolation

We provide four training config files for the four training setups reported in our paper. The training config is set for 4xH100 GPUs. You may need to adjust num_processes and per_device_train_batch_size based on your computation environment.

• Mistral-Base:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml scripts/run_simpo.py training_configs/mistral-7b-base-simpo.yaml

• Mistral-Instruct:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml scripts/run_simpo.py training_configs/mistral-7b-instruct-simpo.yaml

• Llama3-Base:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml scripts/run_simpo.py training_configs/llama-3-8b-base-simpo.yaml

• Llama3-Instruct:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml scripts/run_simpo.py training_configs/llama-3-8b-instruct-simpo.yaml

• Llama3-Instruct v0.2:

ACCELERATE_LOG_LEVEL=info accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml scripts/run_simpo.py training_configs/llama-3-8b-instruct-simpo-v2.yaml

We follow the official implementation for evaluation on AlpacaEval 2, Arena-Hard, and MT-Bench, as follows (more details can be found under the eval directory):

• AlpacaEval 2: Please refer to the AlpacaEval repo for evaluation.

• Arena-Hard: Please refer to to the Arena-Hard-Auto repo for evaluation.

• MT-Bench: Please refer to the FastChat repo for evaluation.

If you have any questions related to the code or the paper, feel free to email Yu ([email protected]). If you encounter any problems when using the code, or want to report a bug, feel free to open an issue! Please try to specify the problem with details so we can help you better and quicker!

Please cite our paper if you find the repo helpful in your work:

@article{meng2024simpo,
  title={{SimPO}: Simple Preference Optimization with a Reference-Free Reward},
  author={Meng, Yu and Xia, Mengzhou and Chen, Danqi},
  journal={arXiv preprint arXiv:2405.14734},
  year={2024}
}