File size: 12,758 Bytes
f39a6b6 8881dd4 f39a6b6 26db919 f39a6b6 f6bdb40 86323c8 f39a6b6 7a8c806 9812423 f39a6b6 97bc38d f39a6b6 eb2676d 86323c8 eb2676d f39a6b6 8881dd4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 |
---
license: llama3
model-index:
- name: ArmoRM-Llama3-8B-v0.1
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: IFEval (0-Shot)
type: HuggingFaceH4/ifeval
args:
num_few_shot: 0
metrics:
- type: inst_level_strict_acc and prompt_level_strict_acc
value: 18.97
name: strict accuracy
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: BBH (3-Shot)
type: BBH
args:
num_few_shot: 3
metrics:
- type: acc_norm
value: 1.75
name: normalized accuracy
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MATH Lvl 5 (4-Shot)
type: hendrycks/competition_math
args:
num_few_shot: 4
metrics:
- type: exact_match
value: 0.0
name: exact match
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GPQA (0-shot)
type: Idavidrein/gpqa
args:
num_few_shot: 0
metrics:
- type: acc_norm
value: 0.0
name: acc_norm
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MuSR (0-shot)
type: TAUR-Lab/MuSR
args:
num_few_shot: 0
metrics:
- type: acc_norm
value: 6.65
name: acc_norm
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU-PRO (5-shot)
type: TIGER-Lab/MMLU-Pro
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 0.87
name: accuracy
source:
url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=RLHFlow/ArmoRM-Llama3-8B-v0.1
name: Open LLM Leaderboard
---
# Absolute-Rating Multi-Objective Reward Model (ArmoRM) with Mixture-of-Experts (MoE) Aggregation of Reward Objectives
+ **Authors** (* indicates equal contribution)
[Haoxiang Wang*](https://haoxiang-wang.github.io/), [Wei Xiong*](https://weixiongust.github.io/WeiXiongUST/index.html), [Tengyang Xie](https://tengyangxie.github.io/), [Han Zhao](https://hanzhaoml.github.io/), [Tong Zhang](https://tongzhang-ml.org/)
+ **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](https://huggingface.co/RLHFlow/ArmoRM-Llama3-8B-v0.1)
+ Finetuned from model: [FsfairX-LLaMA3-RM-v0.1](https://huggingface.co/sfairXC/FsfairX-LLaMA3-RM-v0.1)
- **Code Repository:** https://github.com/RLHFlow/RLHF-Reward-Modeling/
+ **Architecture**
<p align="center">
<img width="800" alt="image" src="https://github.com/RLHFlow/RLHFlow.github.io/blob/main/assets/ArmoRM-MoE.png?raw=true">
</p>
## RewardBench LeaderBoard
| 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](https://huggingface.co/RLHFlow/pair-preference-model-LLaMA3-8B)| Llama-3 8B | [SliC-HF](https://arxiv.org/abs/2305.10425) | 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](https://huggingface.co/sfairXC/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](https://huggingface.co/Nexusflow/Starling-RM-34B) | Yi-34B | Bradley-Terry | 81.4 | 96.9 | 57.2 | 88.2 | 88.5 | 71.4 |
## Demo Code
```python
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
```python
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)
```
## Citation
If you find this work useful for your research, please consider citing:
```
@inproceedings{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},
booktitle={EMNLP},
year={2024}
}
@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](https://arxiv.org/abs/2402.18571)", 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.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_RLHFlow__ArmoRM-Llama3-8B-v0.1)
| Metric |Value|
|-------------------|----:|
|Avg. | 4.71|
|IFEval (0-Shot) |18.97|
|BBH (3-Shot) | 1.75|
|MATH Lvl 5 (4-Shot)| 0.00|
|GPQA (0-shot) | 0.00|
|MuSR (0-shot) | 6.65|
|MMLU-PRO (5-shot) | 0.87|
|