Pairwise Reward Model (PairRM) takes an instruction and a pair of output candidates as the input,
and output a score for each candidate to measure their relative quality.
PairRM can be used to (re-)rank a list of candidate outputs and thus can be used an LLM evaluator to efficiently assess the quality of LLMs in local environment.
PairRM can also be used to enhance the decoding by best-of-n sampling (i.e., reranking N sampled outputs).
Apart from that, one can also use PairRM to further align instruction-tuned LLMs with RLHF methods.
Unlike the other RMs that encode and score each candidate respectively,
PairRM takes a pair of candidates and compares them side-by-side to indentify the subtle differences between them.
Also, PairRM is based on microsoft/deberta-v3-large, and thus it is super efficient: 0.4B.
We trained PairRM on a diverse collection of six human-preference datasets (see more here).
PairRM is part of the LLM-Blender project (ACL 2023). Please see our paper above to know more.
Use Case 1: Comparing/Ranking output candidates given an instruction
Ranking a list candidate responses
python
1inputs =["hello, how are you!","I love you!"]2candidates_texts =[["get out!","hi! I am fine, thanks!","bye!"],3["I love you too!","I hate you!","Thanks! You're a good guy!"]]4ranks = blender.rank(inputs, candidates_texts, return_scores=False, batch_size=1)5# ranks is a list of ranks6# ranks[i][j] represents the ranks of candidate-j for input-i7"""
8ranks -->
9array([[3, 1, 2], # it means "hi! I am fine, thanks!" ranks the 1st, "bye" ranks the 2nd, and "get out!" ranks the 3rd.
10 [1, 3, 2]], # it means "I love you too"! ranks the the 1st, and "I hate you!" ranks the 3rd.
11 dtype=int32)
1213"""
Directly comparing two candidate responses
python
1inputs =["hello!","I love you!"]2candidates_A =["hi!","I hate you!"]3candidates_B =["f**k off!","I love you, too!"]4comparison_results = blender.compare(inputs, candidates_A, candidates_B)5# comparison_results is a list of bool, where comparison_results[i] denotes6# whether candidates_A[i] is better than candidates_B[i] for inputs[i]7# Example: comparison_results[0]--> True
Comparing two multi-turn conversations.
python
1conv1 =[2{3"content":"hello",4"role":"USER"5},6{7"content":"[assistant1‘s response 1]",8"role":"ASSISTANT"9},10...11]12conv2 =[13{14"content":"hello",15"role":"USER"16},17{18"content":"[assistant2's response 1]",19"role":"ASSISTANT"20},21...22]23comparison_results = blender.compare_conversations([conv1],[conv2])24# comparison_results is a list of bool, where each element denotes whether all the responses in conv1 together is better than that of conv2
Use Case 2: Best-of-n Sampling (Decoding Enhancment)
Best-of-n Sampling, aka, rejection sampling, is a strategy to enhance the response quality by selecting the one that was ranked highest by the reward model
(see more in OpenAI WebGPT section 3.2 and OpenAI Blog).
Best-of-n sampling with PairRM is a very easy way to imporve your LLMs with only a few changes of your inference code:
python
1# loading models 2import llm_blender
3from transformers import AutoTokenizer, AutoModelForCausalLM
4tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")5model = AutoModelForCausalLM.from_pretrained("HuggingFaceH4/zephyr-7b-beta", device_map="auto")6system_message ={"role":"system","content":"You are a friendly chatbot."}78# formatting your inputs 9inputs =["can you tell me a joke about OpenAI?"]10messages =[[system_message,{"role":"user","content": _input}]for _input in inputs]11prompts =[tokenizer.apply_chat_template(m, tokenize=False, add_generation_prompt=True)for m in messages]1213# Conventional generation method 14input_ids = tokenizer(prompts[0], return_tensors="pt").input_ids
15sampled_outputs = model.generate(input_ids, do_sample=True, top_k=50, top_p=0.95, num_return_sequences=1)16print(tokenizer.decode(sampled_outputs[0][len(input_ids[0]):], skip_special_tokens=False))17# --> The output could be a bad case such as a very short one, e.g., `Sure` 1819# PairRM for best-of-n sampling 20blender = llm_blender.Blender()21blender.loadranker("llm-blender/PairRM")# load ranker checkpoint22outputs = blender.best_of_n_generate(model, tokenizer, prompts, n=10)2324print("### Prompt:\n", prompts[0])25print("### best-of-n generations:\n", outputs[0])26# --> The output will be much more stable and consistently better than single sampling, for example: 27"""
28Sure, here's a joke about OpenAI:
2930Why did OpenAI decide to hire a mime as their new AI researcher?
3132Because they wanted someone who could communicate complex ideas without making a sound!
3334(Note: This is a joke, not a reflection of OpenAI's actual hiring practices.)
35"""
Use case 3: RLHF
PairRM has been trained on various high-quality and large-scale datasets with human preference annotations
and shown great correlation with human preferences with an extremely small model size (0.4B),
approching the performance of GPT-4.
PairRM will better help the future alignment of LLMs in a more efficient and effective way.
With a blender.compare() function, you can apply PairRM to popular RLHF toolkits such as trl.
🔥 Check more details on our example jupyter notebook usage: blender_usage.ipynb
PairRM has been trained on various high-quality and large-scale dataset with human preference annotations and exhibits great correlation with human preferences
with an extremly small model size (0.4B), approching the performance of GPT-4.
All following results are reported as pairwise comparison accuracies (agreements).
Auto-J Pairwise test data performance
Model
Summ
Exam
Code
Rewriting
Crea W
Func W
Comm
NLP
Overall
Closed -source Models
ChatGPT
33.3
40.3
36.6
31.6
48.2
40.4
47.6
45.8
42.7
Claude -2
30.6
36.1
41.7
34.2
48.1
42.5
40.6
48.5
42.4
GPT -4
59.7
51.4
69.2
58.3
66.7
60.4
58.3
65.2
61.9
Open -source Models
SteamSHP
33.3
29.2
26.7
33.3
40.7
31.3
51.4
51.9
40.6
PandaLM
29.2
33.3
31.7
23.3
43.5
32.9
44.8
48.9
38.9
LLaMA -2-Chat -13B
20.8
27.8
19.2
20
31.5
27.5
35.8
31.8
29
Vicuna -13B-v1.5
30.6
23.6
35
28.3
36.1
37.5
45.5
39.8
37.3
WizardLM -13B-v1.2
22.2
20.8
32.5
19.2
28.7
25.4
29.2
33
27.8
LLAMA -2-chat -70B
34.7
33.3
36.7
35.8
51.4
54.2
47.2
47.7
45.9
AUTO -J (13b)
45.8
38.9
59.2
47.5
54.6
57.1
58
57.6
54.8
UltraRM (13b)
56.94
43.06
55.0
53.33
67.13
64.17
56.25
59.85
59.85
PairRM (0.4b)
56.94
52.78
58.33
55.83
61.57
59.17
57.64
62.5
59.05
HHH-Alignment and MT-bench human judgements
Evaluator LM
HHH ALIGNMENT
MT BENCH HUMAN JUDG .
Help .
Harm .
Hon .
Other
Total Avg .
Human Preference
RANDOM
50
50
50
50
50
34.26
STANFORDNLP REWARD MODEL
69.49
60.34
52.46
51.16
58.82
44.79
ALMOST REWARD MODEL
74.58
67.24
78.69
86.05
76.02
49.9
LLAMA2 -CHAT 7B
66.1
81.03
70.49
74.42
72.85
51.78
LLAMA2 -CHAT 13B
74.58
87.93
55.74
79.07
73.76
52.34
LLAMA2 -CHAT 70B
66.1
89.66
67.21
74.42
74.21
53.67
LLAMA2 -CHAT 13B+COARSE .
68.74
68.97
65.57
67.44
67.42
46.89
GPT -3.5-TURBO -0613
76.27
87.93
67.21
86.05
78.73
57.12
PROMETHEUS 7B
69.49
84.48
78.69
90.7
80.09
55.14
PROMETHEUS 13B
81.36
82.76
75.41
76.74
79.19
57.72
UltraRM (13B)
86.44
79.31
81.97
88.37
83.71
56
PairRM (0.4B)
84.75
84.48
80.33
90.7
84.62
59
GPT -4-0613
91.53
93.1
85.25
83.72
88.69
63.87
While PairRM is a extremely small model (0.4B) based on deberta, the pairwise comparison aggrement performance approches GPT-4's performance!
Two reasons to attribute:
Our PairRM specically designed model arch for pairwise comparison through bidirectional attention (See LLM-blender paper for more details)
The high-quality and large-scale human preference annotation data it was train on (see training dataset list on this hugging face page)
Citation & Credits
If you are using PairRM in your research, please cite LLM-blender.
bibtex
1@inproceedings{llm-blender-2023,
2 title = "LLM-Blender: Ensembling Large Language Models with Pairwise Comparison and Generative Fusion",
3 author = "Jiang, Dongfu and Ren, Xiang and Lin, Bill Yuchen",
4 booktitle = "Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics (ACL 2023)",
5 year = "2023"
6}
7
