Hugging Face's logo

Datasets:
dongboklee
/
dORM-14B-test

Modalities:
Text
Formats:
json
Size:
1K - 10K
Libraries:
Datasets
pandas
Dataset card Data Studio Files
xet
 Community
2

Enhance dataset card: add paper, code, task category, language, abstract, assets, and citation

#2
by nielsr HF Staff - opened
base: refs/heads/main
←
from: refs/pr/2
Discussion Files changed
+101
-15
Files changed (1) hide show
  1. README.md +101 -15
README.md CHANGED
@@ -3,35 +3,50 @@ configs:
3
  - config_name: default
4
  data_files:
5
  - split: law
6
- path: "law.json"
7
  - split: psychology
8
- path: "psychology.json"
9
  - split: chemistry
10
- path: "chemistry.json"
11
  - split: biology
12
- path: "biology.json"
13
  - split: physics
14
- path: "physics.json"
15
  - split: history
16
- path: "history.json"
17
  - split: economics
18
- path: "economics.json"
19
  - split: math
20
- path: "math.json"
21
  - split: business
22
- path: "business.json"
23
  - split: philosophy
24
- path: "philosophy.json"
25
  - split: health
26
- path: "health.json"
27
  - split: engineering
28
- path: "engineering.json"
29
  - split: computer_science
30
- path: "computer_science.json"
31
  - split: other
32
- path: "other.json"
 
 
 
 
33
  ---
34
- # Reward of [test](https://huggingface.co/datasets/dongboklee/test) split extracted by dORM-14B: dORM-14B-test
 
 
 
 
 
 
 
 
 
 
 
35
  ## Usage
36
  ```python
37
  from datasets import load_dataset
@@ -42,3 +57,74 @@ dataset = load_dataset("dongboklee/dORM-14B-test")
42
  # Load specific domain
43
  law_dataset = load_dataset("dongboklee/dORM-14B-test", split="law")
44
  ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  - config_name: default
4
  data_files:
5
  - split: law
6
+ path: law.json
7
  - split: psychology
8
+ path: psychology.json
9
  - split: chemistry
10
+ path: chemistry.json
11
  - split: biology
12
+ path: biology.json
13
  - split: physics
14
+ path: physics.json
15
  - split: history
16
+ path: history.json
17
  - split: economics
18
+ path: economics.json
19
  - split: math
20
+ path: math.json
21
  - split: business
22
+ path: business.json
23
  - split: philosophy
24
+ path: philosophy.json
25
  - split: health
26
+ path: health.json
27
  - split: engineering
28
+ path: engineering.json
29
  - split: computer_science
30
+ path: computer_science.json
31
  - split: other
32
+ path: other.json
33
+ task_categories:
34
+ - text-ranking
35
+ language:
36
+ - en
37
  ---
38
+
39
+ # dORM-14B-test Dataset
40
+
41
+ This dataset contains reward scores for the `test` split, extracted by the `dORM-14B` model, as part of the research presented in the paper "[Rethinking Reward Models for Multi-Domain Test-Time Scaling](https://huggingface.co/papers/2510.00492)".
42
+
43
+ **Paper:** [Rethinking Reward Models for Multi-Domain Test-Time Scaling](https://huggingface.co/papers/2510.00492)
44
+ **Code:** https://github.com/db-Lee/Multi-RM
45
+
46
+ ## Abstract of the Paper
47
+
48
+ The reliability of large language models (LLMs) during test-time scaling is often assessed with \emph{external verifiers} or \emph{reward models} that distinguish correct reasoning from flawed logic. Prior work generally assumes that process reward models (PRMs), which score every intermediate reasoning step, outperform outcome reward models (ORMs) that assess only the final answer. This view is based mainly on evidence from narrow, math-adjacent domains. We present the first unified evaluation of four reward model variants, discriminative ORM and PRM (\DisORM, \DisPRM) and generative ORM and PRM (\GenORM, \GenPRM), across 14 diverse domains. Contrary to conventional wisdom, we find that (i) \DisORM performs on par with \DisPRM, (ii) \GenPRM is not competitive, and (iii) overall, \GenORM is the most robust, yielding significant and consistent gains across every tested domain. We attribute this to PRM-style stepwise scoring, which inherits label noise from LLM auto-labeling and has difficulty evaluating long reasoning trajectories, including those involving self-correcting reasoning. Our theoretical analysis shows that step-wise aggregation compounds errors as reasoning length grows, and our empirical observations confirm this effect. These findings challenge the prevailing assumption that fine-grained supervision is always better and support generative outcome verification for multi-domain deployment. We publicly release our code, datasets, and checkpoints at \href{ this https URL }{\underline{\small\texttt{ this https URL }}} to facilitate future research in multi-domain settings.
49
+
50
  ## Usage
51
  ```python
52
  from datasets import load_dataset
 
57
  # Load specific domain
58
  law_dataset = load_dataset("dongboklee/dORM-14B-test", split="law")
59
  ```
60
+
61
+ ## Assets
62
+
63
+ Please find the assets of this repo below, including training and test datasets, model checkpoints, and rewards obtained by the four reward model variants.
64
+
65
+ ### Datasets
66
+
67
+ | Name | Description |
68
+ |:---|:---|
69
+ | [train](https://huggingface.co/datasets/dongboklee/train) | multi-domain training dataset for dORM/dPRM (mostly adapted from [VersaPRM](https://github.com/UW-Madison-Lee-Lab/VersaPRM)). |
70
+ | [train_gORM](https://huggingface.co/datasets/dongboklee/train_gORM) | multi-domain training dataset for gORM generated by [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B). |
71
+ | [train_gPRM](https://huggingface.co/datasets/dongboklee/train_gPRM) | multi-domain training dataset for gPRM generated by [QwQ-32B](https://huggingface.co/Qwen/QwQ-32B). |
72
+ | [test](https://huggingface.co/datasets/dongboklee/test) | multi-domain test dataset with CoTs (N=128) generated by [Llama3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) (mostly adapted from [VersaPRM](https://github.com/UW-Madison-Lee-Lab/VersaPRM)). |
73
+ | [test_smollm](https://huggingface.co/datasets/dongboklee/test_smollm) | multi-domain test dataset with CoTs (N=16) generated by [SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B). |
74
+ | [test_qwen](https://huggingface.co/datasets/dongboklee/test_qwen) | multi-domain test dataset with CoTs (N=16) generated by [Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct). |
75
+ | [test_gemma](https://huggingface.co/datasets/dongboklee/test_gemma) | multi-domain test dataset with CoTs (N=16) generated by [gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it). |
76
+ | [test_llama](https://huggingface.co/datasets/dongboklee/test_llama) | multi-domain test dataset with CoTs (N=16) generated by [Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct). |
77
+
78
+
79
+ ### Model Checkpoints
80
+
81
+ | Name | Backbone | Trained On | LoRA-merged version |
82
+ |:---|:---|:---|:---|
83
+ | [dORM-14B](https://huggingface.co/dongboklee/dORM-14B) | [14B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) | [train](https://huggingface.co/datasets/dongboklee/train) | — |
84
+ | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [14B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) | [train](https://huggingface.co/datasets/dongboklee/train) | — |
85
+ | [gORM-14B](https://huggingface.co/datasets/dongboklee/gORM-14B) | [14B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) | [train_gORM](https://huggingface.co/datasets/dongboklee/train_gORM) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) |
86
+ | [gPRM-14B](https://huggingface.co/datasets/dongboklee/gPRM-14B) | [14B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) | [train_gPRM](https://huggingface.co/datasets/dongboklee/train_gPRM) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) |
87
+ | [dORM-8B](https://huggingface.co/datasets/dongboklee/dORM-8B) | [8B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) | [train](https://huggingface.co/datasets/dongboklee/train) | — |
88
+ | [dPRM-8B](https://huggingface.co/datasets/dongboklee/dPRM-8B) | [8B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) | [train](https://huggingface.co/datasets/dongboklee/train) | — |
89
+ | [gORM-8B](https://huggingface.co/datasets/dongboklee/gORM-8B) | [8B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) | [train_gORM](https://huggingface.co/datasets/dongboklee/train_gORM) | [gORM-8B-merged](https://huggingface.co/datasets/dongboklee/gORM-8B-merged) |
90
+ | [gPRM-8B](https://huggingface.co/datasets/dongboklee/gPRM-8B) | [8B backbone](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) | [train_gPRM](https://huggingface.co/datasets/dongboklee/train_gPRM) | [gPRM-8B-merged](https://huggingface.co/datasets/dongboklee/gPRM-8B-merged) |
91
+
92
+
93
+ ### Rewards on Datasets by Model
94
+
95
+ | Name | Model | Dataset |
96
+ |:---|:---|:---|
97
+ | [dORM-14B-test](https://huggingface.co/datasets/dongboklee/dORM-14B-test) | [dORM-14B](https://huggingface.co/datasets/dongboklee/dORM-14B) | [test](https://huggingface.co/datasets/dongboklee/test) |
98
+ | [dORM-8B-test](https://huggingface.co/datasets/dongboklee/dORM-8B-test) | [dORM-8B](https://huggingface.co/datasets/dongboklee/dORM-8B) | [test](https://huggingface.co/datasets/dongboklee/test) |
99
+ | [dORM-14B-test_smollm](https://huggingface.co/datasets/dongboklee/dORM-14B-test_smollm) | [dORM-14B](https://huggingface.co/datasets/dongboklee/dORM-14B) | [test_smollm](https://huggingface.co/datasets/dongboklee/test_smollm) |
100
+ | [dORM-14B-test_qwen](https://huggingface.co/datasets/dongboklee/dORM-14B-test_qwen) | [dORM-14B](https://huggingface.co/datasets/dongboklee/dORM-14B) | [test_qwen](https://huggingface.co/datasets/dongboklee/test_qwen) |
101
+ | [dORM-14B-test_gemma](https://huggingface.co/datasets/dongboklee/dORM-14B-test_gemma) | [dORM-14B](https://huggingface.co/datasets/dongboklee/dORM-14B) | [test_gemma](https://huggingface.co/datasets/dongboklee/test_gemma) |
102
+ | [dORM-14B-test_llama](https://huggingface.co/datasets/dongboklee/dORM-14B-test_llama) | [dORM-14B](https://huggingface.co/datasets/dongboklee/dORM-14B) | [test_llama](https://huggingface.co/datasets/dongboklee/test_llama) |
103
+ | [dPRM-14B-test](https://huggingface.co/datasets/dongboklee/dPRM-14B-test) | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [test](https://huggingface.co/datasets/dongboklee/test) |
104
+ | [dPRM-8B-test](https://huggingface.co/datasets/dongboklee/dPRM-8B-test) | [dPRM-8B](https://huggingface.co/datasets/dongboklee/dPRM-8B) | [test](https://huggingface.co/datasets/dongboklee/test) |
105
+ | [dPRM-14B-test_smollm](https://huggingface.co/datasets/dongboklee/dPRM-14B-test_smollm) | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [test_smollm](https://huggingface.co/datasets/dongboklee/test_smollm) |
106
+ | [dPRM-14B-test_qwen](https://huggingface.co/datasets/dongboklee/dPRM-14B-test_qwen) | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [test_qwen](https://huggingface.co/datasets/dongboklee/test_qwen) |
107
+ | [dPRM-14B-test_gemma](https://huggingface.co/datasets/dongboklee/dPRM-14B-test_gemma) | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [test_gemma](https://huggingface.co/datasets/dongboklee/test_gemma) |
108
+ | [dPRM-14B-test_llama](https://huggingface.co/datasets/dongboklee/dPRM-14B-test_llama) | [dPRM-14B](https://huggingface.co/datasets/dongboklee/dPRM-14B) | [test_llama](https://huggingface.co/datasets/dongboklee/test_llama) |
109
+ | [gORM-14B-test](https://huggingface.co/datasets/dongboklee/gORM-14B-test) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) | [test](https://huggingface.co/datasets/dongboklee/test) |
110
+ | [gORM-8B-test](https://huggingface.co/datasets/dongboklee/gORM-8B-test) | [gORM-8B-merged](https://huggingface.co/datasets/dongboklee/gORM-8B-merged) | [test](https://huggingface.co/datasets/dongboklee/test) |
111
+ | [gORM-14B-test_smollm](https://huggingface.co/datasets/dongboklee/gORM-14B-test_smollm) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) | [test_smollm](https://huggingface.co/datasets/dongboklee/test_smollm) |
112
+ | [gORM-14B-test_qwen](https://huggingface.co/datasets/dongboklee/gORM-14B-test_qwen) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) | [test_qwen](https://huggingface.co/datasets/dongboklee/test_qwen) |
113
+ | [gORM-14B-test_gemma](https://huggingface.co/datasets/dongboklee/gORM-14B-test_gemma) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) | [test_gemma](https://huggingface.co/datasets/dongboklee/test_gemma) |
114
+ | [gORM-14B-test_llama](https://huggingface.co/datasets/dongboklee/gORM-14B-test_llama) | [gORM-14B-merged](https://huggingface.co/datasets/dongboklee/gORM-14B-merged) | [test_llama](https://huggingface.co/datasets/dongboklee/test_llama) |
115
+ | [gPRM-14B-test](https://huggingface.co/datasets/dongboklee/gPRM-14B-test) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) | [test](https://huggingface.co/datasets/dongboklee/test) |
116
+ | [gPRM-8B-test](https://huggingface.co/datasets/dongboklee/gPRM-8B-test) | [gPRM-8B-merged](https://huggingface.co/datasets/dongboklee/gPRM-8B-merged) | [test](https://huggingface.co/datasets/dongboklee/test) |
117
+ | [gPRM-14B-test_smollm](https://huggingface.co/datasets/dongboklee/gPRM-14B-test_smollm) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) | [test_smollm](https://huggingface.co/datasets/dongboklee/test_smollm) |
118
+ | [gPRM-14B-test_qwen](https://huggingface.co/datasets/dongboklee/gPRM-14B-test_qwen) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) | [test_qwen](https://huggingface.co/datasets/dongboklee/test_qwen) |
119
+ | [gPRM-14B-test_gemma](https://huggingface.co/datasets/dongboklee/gPRM-14B-test_gemma) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) | [test_gemma](https://huggingface.co/datasets/dongboklee/test_gemma) |
120
+ | [gPRM-14B-test_llama](https://huggingface.co/datasets/dongboklee/gPRM-14B-test_llama) | [gPRM-14B-merged](https://huggingface.co/datasets/dongboklee/gPRM-14B-merged) | [test_llama](https://huggingface.co/datasets/dongboklee/test_llama) |
121
+
122
+ ## Citation
123
+ ```bibtex
124
+ @article{multi-rm,
125
+ title = {Rethinking Reward Models for Multi-Domain Test-Time Scaling},
126
+ author = {Lee, Dong Bok and Lee, Seanie and Park, Sangwoo and Kang, Minki and Baek, Jinheon and Kim, Dongki and Wagner, Dominik and Jin, Jiongdao and Lee, Heejun and Bocklet, Tobias and Wang, Jinyu and Fu, Jingjing and Hwang, Sung Ju and Bian, Jiang and Song, Lei},
127
+ journal = {arXiv preprint arXiv:2510.00492},
128
+ year = {2025}
129
+ }
130
+ ```