armo-rm

Absolute-Rating Multi-Objective Reward Model (ArmoRM) with Mixture-of-Experts (MoE) Aggregation of Reward Objectives

• Authors (* indicates equal contribution)

  Haoxiang Wang*, Wei Xiong*, Tengyang Xie, Han Zhao, Tong Zhang

• Blog: https://rlhflow.github.io/posts/2024-05-29-multi-objective-reward-modeling/

• Tech Report: https://arxiv.org/abs/2406.12845

• Model: ArmoRM-Llama3-8B-v0.1

  • Finetuned from model: FsfairX-LLaMA3-RM-v0.1

• Code Repository: https://github.com/RLHFlow/RLHF-Reward-Modeling/

• Architecture

  

RewardBench LeaderBoard (May 2024)

Model	Base Model	Method	Score	Chat	Chat Hard	Safety	Reasoning	Prior Sets (0.5 weight)
ArmoRM-Llama3-8B-v0.1	Llama-3 8B	ArmoRM + MoE	89.0	96.9	76.8	92.2	97.3	74.3
Cohere May 2024	Unknown	Unknown	88.3	96.4	71.3	92.7	97.7	78.2
pair-preference-model	Llama-3 8B	SliC-HF	85.7	98.3	65.8	89.7	94.7	74.6
GPT-4 Turbo (0125 version)	GPT-4 Turbo	LLM-as-a-Judge	84.3	95.3	74.3	87.2	86.9	70.9
FsfairX-LLaMA3-RM-v0.1	Llama-3 8B	Bradley-Terry	83.6	99.4	65.1	87.8	86.4	74.9
Starling-RM-34B	Yi-34B	Bradley-Terry	81.4	96.9	57.2	88.2	88.5	71.4

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
device = "cuda"
path = "RLHFlow/ArmoRM-Llama3-8B-v0.1"
model = AutoModelForSequenceClassification.from_pretrained(path, device_map=device, 
                               trust_remote_code=True, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(path, use_fast=True)
# We load a random sample from the validation set of the HelpSteer dataset
prompt = 'What are some synonyms for the word "beautiful"?'
response = "Nicely, Beautifully, Handsome, Stunning, Wonderful, Gorgeous, Pretty, Stunning, Elegant"
messages = [{"role": "user", "content": prompt},
           {"role": "assistant", "content": response}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
with torch.no_grad():
   output = model(input_ids)
   # Multi-objective rewards for the response
   multi_obj_rewards = output.rewards.cpu().float() 
   # The gating layer's output is conditioned on the prompt
   gating_output = output.gating_output.cpu().float()
   # The preference score for the response, aggregated from the 
   # multi-objective rewards with the gating layer
   preference_score = output.score.cpu().float()  
# We apply a transformation matrix to the multi-objective rewards
# before multiplying with the gating layer's output. This mainly aims
# at reducing the verbosity bias of the original reward objectives
obj_transform = model.reward_transform_matrix.data.cpu().float()
# The final coefficients assigned to each reward objective
multi_obj_coeffs = gating_output @ obj_transform.T
# The preference score is the linear combination of the multi-objective rewards with
# the multi-objective coefficients, which can be verified by the following assertion
assert torch.isclose(torch.sum(multi_obj_rewards * multi_obj_coeffs, dim=1), preference_score, atol=1e-3) 
# Find the top-K reward objectives with coefficients of the highest magnitude
K = 3
top_obj_dims = torch.argsort(torch.abs(multi_obj_coeffs), dim=1, descending=True,)[:, :K]
top_obj_coeffs = torch.gather(multi_obj_coeffs, dim=1, index=top_obj_dims)

# The attributes of the 19 reward objectives
attributes = ['helpsteer-helpfulness','helpsteer-correctness','helpsteer-coherence',
   'helpsteer-complexity','helpsteer-verbosity','ultrafeedback-overall_score',
   'ultrafeedback-instruction_following', 'ultrafeedback-truthfulness',
   'ultrafeedback-honesty','ultrafeedback-helpfulness','beavertails-is_safe',
   'prometheus-score','argilla-overall_quality','argilla-judge_lm','code-complexity',
   'code-style','code-explanation','code-instruction-following','code-readability']

example_index = 0
for i in range(K):
   attribute = attributes[top_obj_dims[example_index, i].item()]
   coeff = top_obj_coeffs[example_index, i].item()
   print(f"{attribute}: {round(coeff,5)}")
# code-complexity: 0.19922
# helpsteer-verbosity: -0.10864
# ultrafeedback-instruction_following: 0.07861

# The actual rewards of this example from the HelpSteer dataset
# are [3,3,4,2,2] for the five helpsteer objectives: 
# helpfulness, correctness, coherence, complexity, verbosity
# We can linearly transform our predicted rewards to the 
# original reward space to compare with the ground truth
helpsteer_rewards_pred = multi_obj_rewards[0, :5] * 5 - 0.5
print(helpsteer_rewards_pred)
# [2.78125   2.859375  3.484375  1.3847656 1.296875 ]

Easy to use Pipeline

from typing import Dict, List
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer


class ArmoRMPipeline:
    def __init__(self, model_id, device_map="auto", torch_dtype=torch.bfloat16, truncation=True, trust_remote_code=False, max_length=4096):
        self.model = AutoModelForSequenceClassification.from_pretrained(
            model_id,
            device_map=device_map,
            trust_remote_code=trust_remote_code,
            torch_dtype=torch_dtype,
        )
        self.tokenizer = AutoTokenizer.from_pretrained(
            model_id,
            use_fast=True,
        )
        self.truncation = truncation
        self.device = self.model.device
        self.max_length = max_length

    def __call__(self, messages: List[Dict[str, str]]) -> Dict[str, float]:
        """
        messages: OpenAI chat messages to be scored
        Note: no batching since due to length differences, the model will have to pad to the max length which is not efficient
        Returns: a dictionary with the score between 0 and 1
        """
        input_ids = self.tokenizer.apply_chat_template(
            messages,
            return_tensors="pt",
            padding=True,
            truncation=self.truncation,
            max_length=self.max_length,
        ).to(self.device)
        with torch.no_grad():
            output = self.model(input_ids)
            score = output.score.float().item()
        return {"score": score}

# Create Reward Model Pipeline 
prompt = 'What are some synonyms for the word "beautiful"?'
rm = ArmoRMPipeline("RLHFlow/ArmoRM-Llama3-8B-v0.1", trust_remote_code=True)
# score the messages
response1 = 'Nicely, Beautifully, Handsome, Stunning, Wonderful, Gorgeous, Pretty, Stunning, Elegant'
score1 = rm([{"role": "user", "content": prompt}, {"role": "assistant", "content": response1}])
print(score1)

response2 = '''Certainly! Here are some synonyms for the word "beautiful":

1. Gorgeous
2. Lovely
3. Stunning
4. Attractive
5. Pretty
6. Elegant
7. Exquisite
8. Handsome
9. Charming
10. Alluring
11. Radiant
12. Magnificent
13. Graceful
14. Enchanting
15. Dazzling

These synonyms can be used in various contexts to convey the idea of beauty.'''
score2 = rm([{"role": "user", "content": prompt}, {"role": "assistant", "content": response2}])
print(score2)

response3 = 'Sorry i cannot answer this.'
score3 = rm([{"role": "user", "content": prompt}, {"role": "assistant", "content": response3}])
print(score3)

Training

ArmoRM has two training stages: 1) Multi-objective Reward Learning and 2) Mixture-of-Experts Gating Network Learning.

Multi-objective Reward Learning

This stage involves training a multi-objective reward model by linear probing on top of an existing reward model. The process includes:

1. Data Preparation (stage-1_prepare.py):

   • Load a multi-objective dataset and extract embeddings from an existing reward model for each example.
   • Save embeddings and labels for further processing.
   • Example Command:

     python stage-1_prepare.py --model_path sfairXC/FsfairX-LLaMA3-RM-v0.1 \
                               --dataset_path RLHFlow/ArmoRM-Multi-Objective-Data-v0.1 \
                               --n_shards 1 --shard_idx 1 --device 0
     

   • The dataset sharding (specified by n_shards and shard_idx) is optional but can be used for parallel processing.

2. Training (stage-1_train.py):

   • Perform multi-objective linear regression on the prepared embeddings, with sklearn's Ridge regression using different L2 regularization strengths for each attribute. Note: this training doesn't require a GPU.
   • Select the best regularization strength based on validation loss.
   • Save the final regression model weights in PyTorch format.
   • Example Command:

     python stage-1_train.py --model_path sfairXC/FsfairX-LLaMA3-RM-v0.1 \
                             --dataset_path RLHFlow/ArmoRM-Multi-Objective-Data-v0.1
     

Key aspects:

• Uses linear probing (training only the new linear layer while keeping transformer layers frozen).
• Deals with missing labels in the merged dataset by ignoring them during loss computation.

Mixture-of-Experts Gating Network Learning

This stage involves training a gating network to aggregate the multi-objective rewards based on the context. The process includes:

1. Data Preparation (stage-2_prepare.py):

   • Load a binary preference dataset and prepare embeddings for prompts and responses using the same reward model as Stage 1.
   • Save embeddings for both chosen and rejected responses in each preference pair.
   • Example Commands:

     python stage-2_prepare.py --model_path sfairXC/FsfairX-LLaMA3-RM-v0.1 --model_family llama3 \
                               --dataset_path RLHFlow/pair_data_v2_80K_wsafety \
                               --dataset_split train --n_shards 1 --shard_idx 1 --device 0
     

   Note:

   • You can optionally run this script with dataset_path referring to a reference dataset, which can be used to mitigate the verbosity bias in the following training step (e.g., RLHFlow/UltraFeedback-preference-standard, which is the reference dataset used in our paper).
   • You can also run this script with --dataset_path allenai/reward-bench --dataset_split filtered to prepare the embeddings for the RewardBench dataset, which can be used in the following training script for evaluation.

2. Training (stage-2_train.py):

   • Verbosity Debiasing: Debiases the multi-objective rewards by applying a transformation matrix to the rewards. The coefficients of the verbosity debiasing are found by a grid search method.
   • Gating Network Learning: Train a gating network (MLP) to aggregate multi-objective rewards based on the prompt.
   • Loss Function: Use Bradley-Terry loss for training on pairwise preference data.
   • Example Command:

     python stage-2_train.py --model_path sfairXC/FsfairX-LLaMA3-RM-v0.1 --model_family llama3 \
                             --multi_objective_dataset RLHFlow/ArmoRM-Multi-Objective-Data-v0.1 \
                             --preference_dataset RLHFlow/pair_data_v2_80K_wsafety \
                             --reference_dataset RLHFlow/UltraFeedback-preference-standard \
                             --eval_reward_bench --device 0
     

     Note:
     • If you do not specify --reference_dataset, the preference dataset will be used as the reference dataset for verbosity debiasing.
     • If you have run stage-2_prepare.py with the --dataset_path allenai/reward-bench --dataset_split filtered, you can run this training script with --eval_reward_bench` to evaluate the model on the RewardBench dataset.

Key aspects:

• Freezes the backbone and regression layer from Stage 1, only training the gating network.
• Conducts verbosity debiasing on the multi-objective rewards.
• Trains the gating network with a Bradley-Terry loss.
• Evaluates performance on the RewardBench benchmark.

Citation

If you find this work useful for your research, please consider citing:

@article{ArmoRM,
      title={Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts}, 
      author={Haoxiang Wang and Wei Xiong and Tengyang Xie and Han Zhao and Tong Zhang},
      journal={arXiv preprint arXiv:2406.12845},
}

@inproceedings{wang2024arithmetic,
      title={Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards}, 
      author={Haoxiang Wang and Yong Lin and Wei Xiong and Rui Yang and Shizhe Diao and Shuang Qiu and Han Zhao and Tong Zhang},
      year={2024},
      booktitle={ACL},
}

The second entry, "Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards", is another recent work of ours that trained a multi-objective reward model and adopted it for LLM alignment, which motivated us to develop the current work.