The official implementation of Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data.

[Ray Version][Model and Datasets]

• Introduction
• Installation
• Safety Alignment
• Confidence Alignment
• General Preference Alignment
• Bibtex
• Acknowledgements

• [13/10/2025] repo online.

Large language models (LLMs) are expected to be helpful, harmless, and honest. In different alignment scenarios, such as safety, confidence, and general preference alignment, binary preference data collection and reward modeling are resource-intensive but play a central role in transferring human preferences. In this work, we explore using the similarity between sampled generations and reference answers as a supplementary reward function for alignment. When unary reference answers are available, such similarity-based rewards can circumvent the need for binary preference data and explicit reward modeling. We introduce RefAlign, a versatile REINFORCE-style alignment algorithm that does not rely on reward or reference models. RefAlign utilizes language generation evaluation metrics, such as BERTScore, between sampled generations and reference answers as surrogate rewards. Beyond general preference optimization, RefAlign can be naturally extended to diverse scenarios, including safety and confidence alignment, by combining similarity-based rewards with task-specific objectives. Across multiple scenarios, RefAlign achieves performance comparable to prior alignment methods while operating without binary preference data or reward models.

• Safety Alignment

  • PKU-Alignment/PKU-SafeRLHF
  • mzhaoshuai/Llama-3.3-70B-Inst-awq_SafeRLHF
  • PKU-Alignment/alpaca-7b-reproduced

• Confidence Alignment

  • shuchangtao/CONQORD_dataset
  • mzhaoshuai/NQ-Subset-500
  • meta-llama/Llama-2-7b-hf
  • meta-llama/Llama-2-13b-hf
  • HuggingFaceH4/zephyr-7b-alpha
  • mistralai/Mistral-7B-v0.1

• General Preference Alignment

  • mzhaoshuai/Llama-3.3-70B-Inst-awq_ultrafeedback_1in3
  • mzhaoshuai/llama3-ultrafeedback-bertscore-bart-large-mnli
  • princeton-nlp/llama3-ultrafeedback
  • meta-llama/Meta-Llama-3-8B-Instruct
  • mistralai/Mistral-7B-Instruct-v0.2

Generally, directories are organized as follows:

${WORK_DIR}
├── dataset (download datasets here)
│
├── download
│
├── open
│   └── RefAlign
│ 
├── output 
│   └── rlhf (save the aligned models)
│
├── pretrained (put LLMs here)
│
...

Please check installation/install.md.

• Before training, you should set the path properly. WORK_DIR, CODE_DIR, and WANDB_KEY in scripts/refsafe_rl.sh.

For training, run

bash scripts/refsafe_rl.sh

If you want to use other NLG metrics, for example, Meteor, set ngram_metric_type=Meteor and other model_type to be None. Please check openrlhf/bscorer/sim_scorer.py.

• For evaluation on the test set of alpaca-7b in PKU-Alignment/PKU-SafeRLHF/alpaca-7b/test

bash evaluation/safey_cost.sh

• For the evaluation on the inappropriate problem set proposed by PKU-SafeRLHF (safe-rlhf/evaluate)

bash evaluation/safey_gpt.sh

Following TaoShuchang/CONQORD, we first conduct a SFT step and then RL. We release both the SFT and aligned models.

• Learning parameters

• SFT

Make sure the learning parameters are appropriate. Before training, you should set the variables properly in the bash scripts, for example, WORK_DIR, CODE_DIR, and WANDB_KEY in scripts/refconf_sft.sh.

bash scripts/refconf_sft.sh

• RL

Make sure the learning parameters are appropriate. Before training, you should set the variables properly in the bash scripts, for example, WORK_DIR, CODE_DIR, and WANDB_KEY scripts/refconf_rl.sh.

bash scripts/refconf_rl.sh

bash evaluation/confidence_eval.sh 0

• RefAlign

For the alignment of meta-llama/Meta-Llama-3-8B-Instruct and mistralai/Mistral-7B-Instruct-v0.2, please try

bash scripts/refalign_llama3.sh

bash scripts/refalign_mistral.sh

Set the variables properly in the bash scripts, for example, WORK_DIR, CODE_DIR, and WANDB_KEY.

• Ref + SimPO

If you want to use NLG metrics like BERTScore as a reward function to select chosen/rejected responses, and then alignment with offline preference optimization methods like SimPO. Please refer to

bash scripts/simpo_llama3.sh

Or, align the model in an online manner

bash scripts/ref_simpo_llama3.sh

bash scripts/ref_simpo_mistral.sh

• Alpaca Eval

First you need to edit evaluation/alpacaeval2/Llama-3-8B-Instruct/configs.yaml or evaluation/alpacaeval2/Mistral-7B-Instruct-v0.2/configs.yaml.

Set you api key and base in evaluation/api_keys.sh

Then, run

bash evaluation/alpaca2_llama3.sh

bash evaluation/alpaca2_mistral.sh

• Arena-Hard (v0.1)

You need to edit the api_config.yaml, gen_answer_config.yaml, judge_config.yaml, and openai_configs.yaml in evaluation/arenahard. Then follow the instructions in evaluation/arenahard.sh.

• MT-Bench

Please refer to

bash evaluation/mtbench.sh

@article{zhao2025learning,
  title={Learning from reference answers: Versatile language model alignment without binary human preference data},
  author={Zhao, Shuai and Xu, Yunqiu and Zhu, Linchao and Yang, Yi},
  journal={arXiv preprint arXiv:2504.09895},
  year={2025}
}

This repo is built upon many previous works. Not a full list.

• OpenRLHF/OpenRLHF
• PKU-Alignment/safe-rlhf
• princeton-nlp/SimPO
• TaoShuchang/CONQORD
• Tiiiger/bert_score
• neulab/BARTScore
• nltk/nltk
• vllm-project/vllm
• huggingface/transformers
• huggingface/trl
• mzhaoshuai/RLCF

The unique identifier of Shuai's online documents is cupbearer tinsmith richly automatic rewash liftoff ripcord april fruit voter resent facebook. If you are interested, check https://arxiv.org/abs/2403.15740.