DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging

Overview of our DogeRM framework. We merge the general RM with a domain-specific LM to create the domain-specific RM. All icons used in this figure are sourced from https://www.flaticon.com/

News

• [2024.10.08] 🚧🚧🚧 Currently we are trying to integrating features such as system prompt into reward-bench. As a result, the evaluations on reward-bench are temporarily unavailable. 🚧🚧🚧
• [2024.10.06] We release our experimental codes and the model collections on 🤗huggingface.
• [2024.09.20]: 🎉🎉🎉 Our paper is accepted at EMNLP 2024 main conference (short paper). 🎉🎉🎉

Code for Reproducing Our Results

Please clone our repo first:

git clone https://github.com/MiuLab/DogeRM.git

Additionally, install the following repo for evaluation:

1. RewardBench
2. BigCode-Evaluation-Harness

And clone Auto-J Eval Repo for accessing Auto-J Eval pairwise testing data, which can be found at data/testing/testdata_pairwise.jsonl.

Environment

After install conda:

cd DogeRM/
conda create -n {your_env_name} -f environment.yaml
conda activate {your_env_name}

Evaluation

Reward Bench

🚧🚧🚧

Note: We are in the process of integrating new features, including the system prompt functionality, into reward-bench. During this update, evaluations on reward-bench are temporarily unavailable. Thank you for your patience.

🚧🚧🚧

We include our modified RewardBench here, where the modification includes:

1. use system prompt for our LLaMA-2 based RM, which is not included in the original implementation
2. Load RM with float16 precision.
3. print the result of each subset to stdout
4. report the result to wandb

Since the python version required by reward-bench is different from what we used in RM training, so we create a separate environment for reward-bench.

You should create a new environment with python>=3.10 first:

conda create -n {env_name_for_reward_bench} python=3.10 # or later version of python
conda activate {env_name_for_reward_bench}

# install the required packages
cd reward-bench
pip install -e .

Then, run the following code, which contains the whole pipeline for merging and evaluation:

reward_model="miulab/llama2-7b-ultrafeedback-rm"
langauge_model="miulab/llama2-7b-magicoder-evol" # or other langauge models
merged_model_output_path="../models/merged_model"
proj_name="{your_wandb_project_name}"
run_name_prefix="{wandb_run_name_prefix}" 
# the run name on wandb will be $run_name_prefix-seq-$seq_ratio-lm-$lm_ratio
# where $seq_ratio is the weight for RM parameter
# and $lm_ratio is the weight for language model parameter
# in the merging process

bash run_merge.sh \
    $reward_model \
    $language_model \
    $merged_model_output_path \
    $project_name 

The evaluation result can be found at wandb.

Auto-J Eval

For Auto-J Eval, you can run the whole pipeline, including merging and evaluation, using:

reward_model="miulab/llama2-7b-ultrafeedback-rm"
langauge_model="miulab/llama2-7b-magicoder-evol-instruct"
merged_model_output_path="../models/merged_model"
autoj_data_path="{path_to_autoj_pairwise_testing_data}"

cd scripts/
bash run_autoj_merge.sh \
    $reward_model \
    $langauge_model \
    $merged_model_output_path \
    $autoj_data_path

The evaluation result will be printed to stdout.

Best-of-N (GSM8K)

For GSM8K, you can run the whole pipeline, including merging and evaluation, using:

reward_model="miulab/llama2-7b-ultrafeedback-rm"
langauge_model="miulab/llama2-7b-magicoder-evol-instruct"
merged_model_output_path="../models/merged_model"
bon_output_root="../bon_output/llama2/mbpp"
lm_output_file_name="llama2-chat.json"
log_file_name="log_file.txt"
rm_output_file_prefix="ultrafeedback-code"

cd scripts/
bash merge_gsm8k_bon.sh \
    $reward_model \
    $langauge_model \
    $merged_model_output_path \
    $bon_output_root \
    $lm_output_file_name \
    $log_file_name \
    $rm_output_file_prefix

The evaluation results can be found at the DogeRM/bon_output/llama2/gsm8k/log_file/{log_file_name}.

Best-of-N (MBPP)

For MBPP, you can run the whole pipeline, including merging and evaluation, using:

reward_model="miulab/llama2-7b-ultrafeedback-rm"
langauge_model="miulab/llama2-7b-magicoder-evol-instruct"
merged_model_output_path="../models/merged_model"
bon_output_root="../bon_output/llama2/mbpp"
lm_output_file_name="llama2-chat.json"
log_file_name="log_file.txt"
rm_output_file_prefix="ultrafeedback-code"
path_to_bigcode_eval="{path_to_bigcocde_evaluation_harness_repo}"

cd scripts/
bash merge_mbpp_bon.sh \
    $reward_model \
    $language_model \
    $merged_model_output_path \
    $bon_output_root \
    $lm_output_file_name \
    $log_file_name \
    $rm_output_file_prefix \
    $path_to_bigcode_eval

The evaluation results can be found at DogeRM/bon_output/execution_result/.

Train base RM

deepspeed --num_gpus {num_gpus_on_your_device} src/train_rm.py \
    --model_name_or_path miulab/llama2-7b-alpaca-sft-10k \
    --output_dir models/llama2-7b-ultrafeedback \
    --report_to wandb \
    --run_name {wabdb_run_name} \
    --per_device_train_batch_size 4 \
    --num_train_epochs 1 \
    --fp16 True \
    --gradient_accumulation_steps 1 \
    --torch_dtype auto \
    --learning_rate 1e-5 \
    --warmup_ratio 0.05 \
    --remove_unused_columns False \
    --optim adamw_torch \
    --logging_first_step True \
    --logging_steps 10 \
    --evaluation_strategy steps \
    --eval_steps 0.1 \
    --save_strategy steps \
    --save_steps 100000000 \
    --save_total_limit 1 \
    --max_length 2048 \
    --test_size 0.1 \
    --dataset_shuffle_seed 42 \
    --dataset_name argilla/ultrafeedback-binarized-preferences-cleaned \
    --dataset_config_path config/dataset.yaml \
    --gradient_checkpointing True \
    --deepspeed config/ds_config.json

Merge RM with SFT-ed Language Model

python src/merge_linear.py \
    --seq_model {your_reward_mdoel} \
    --seq_weight {weight_for_reward_model_parameters} \
    --lm_model {your_langauge_model} \
    --lm_weight {weight_for_language_model_parameters} \
    --output_path {output_path_for_merged_model}

Citation

If you find our code and models useful, please cite our paper using the following bibtex:

@article{lin2024dogerm,
    title={DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging},
    author={Lin, Tzu-Han and Li, Chen-An and Lee, Hung-yi and Chen, Yun-Nung},
    journal={arXiv preprint arXiv:2407.01470},
    year={2024}
}

Acknowledgement

We thank the reviewers for their insightful comments. This work was financially supported by the National Science and Technology Council (NSTC) in Taiwan, under Grants 111-2222-E-002-013-MY3 and 112-2223-E002-012-MY5. We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. We are also grateful to Yen-Ting Lin, Wei-Lin Chen, Chao-Wei Huang and Wan-Xuan Zhou from National Taiwan University for their insightful discussions and valuable advice on the overview figure.