lin99 commited on
Commit
239a5e9
·
verified ·
1 Parent(s): 30b67de

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +156 -0
README.md ADDED
@@ -0,0 +1,156 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ task_categories:
4
+ - question-answering
5
+ - text-generation
6
+ language:
7
+ - en
8
+ tags:
9
+ - benchmark
10
+ - professional
11
+ - expert
12
+ - reasoning
13
+ - law
14
+ - medicine
15
+ - finance
16
+ - cybersecurity
17
+ - engineering
18
+ - llm-evaluation
19
+ - multiple-choice
20
+ size_categories:
21
+ - 1K<n<10K
22
+ ---
23
+
24
+ # ProBench
25
+
26
+ A 4-choice multiple-choice benchmark of **529 expert-level questions** across five professional domains, sourced entirely from Stack Exchange communities (CC-BY-SA 4.0). Designed to evaluate LLMs where standard benchmarks (MMLU, HumanEval, HellaSwag) are now saturated above 90%.
27
+
28
+ ## Why ProBench?
29
+
30
+ | Benchmark | Frontier model score | Status |
31
+ |---|---|---|
32
+ | MMLU | >90% | Saturated |
33
+ | HellaSwag | >95% | Saturated |
34
+ | HumanEval | >90% | Saturated |
35
+ | **ProBench** | TBD | Active |
36
+
37
+ Questions are sourced from real professional practitioners. Correct answers are community-verified (Stack Exchange accepted answers with high vote scores). Distractors are real expert-written text from the same domain — not synthetically generated.
38
+
39
+ ## Dataset Structure
40
+
41
+ ### Splits
42
+
43
+ | Split | Count |
44
+ |---|---|
45
+ | train | 370 |
46
+ | validation | 79 |
47
+ | test | 80 |
48
+ | **Total** | **529** |
49
+
50
+ ### By Domain
51
+
52
+ | Domain | Source Community | Items |
53
+ |---|---|---|
54
+ | Cybersecurity | security.stackexchange.com | 148 |
55
+ | Law | law.stackexchange.com | 121 |
56
+ | Medicine | medicalsciences.stackexchange.com | 63 |
57
+ | Engineering | engineering.stackexchange.com | 113 |
58
+ | Finance | quant.stackexchange.com | 84 |
59
+
60
+ ### Record Format
61
+
62
+ ```json
63
+ {
64
+ "id": "cybersecurity_29988",
65
+ "domain": "cybersecurity",
66
+ "question_title": "What is certificate pinning?",
67
+ "question_body": "I've recently seen ...",
68
+ "question_score": 373,
69
+ "question_tags": ["tls", "certificates", "public-key-infrastructure"],
70
+ "choices": {
71
+ "A": "You can roll your own, but you probably will make a major security...",
72
+ "B": "Software is too complex. This is by far the most important factor...",
73
+ "C": "The known_hosts file lets the client authenticate the server...",
74
+ "D": "Typically certificates are validated by checking the signature hierarchy..."
75
+ },
76
+ "answer": "D",
77
+ "distractor_source": "same_domain_answer_pool",
78
+ "source": "stackexchange",
79
+ "license": "CC-BY-SA 4.0",
80
+ "url": "https://security.stackexchange.com/questions/29988/..."
81
+ }
82
+ ```
83
+
84
+ ## Evaluation
85
+
86
+ Evaluation is exact-match (0 or 1 per question) — fully automated, no human judge required.
87
+
88
+ ```python
89
+ from datasets import load_dataset
90
+
91
+ ds = load_dataset("lin99/ProBench", split="test")
92
+
93
+ def evaluate(model, dataset):
94
+ correct = 0
95
+ for row in dataset:
96
+ prompt = f"""Domain: {row['domain']}
97
+
98
+ Question: {row['question_title']}
99
+
100
+ {row['question_body']}
101
+
102
+ A. {row['choices']['A']}
103
+ B. {row['choices']['B']}
104
+ C. {row['choices']['C']}
105
+ D. {row['choices']['D']}
106
+
107
+ Answer (A/B/C/D):"""
108
+ prediction = model(prompt).strip()[0].upper()
109
+ if prediction == row["answer"]:
110
+ correct += 1
111
+ return correct / len(dataset)
112
+
113
+ accuracy = evaluate(your_model, ds)
114
+ print(f"ProBench accuracy: {accuracy:.1%}")
115
+ ```
116
+
117
+ ## Quality Filters
118
+
119
+ | Filter | Threshold |
120
+ |---|---|
121
+ | Minimum question score | 10 upvotes |
122
+ | Minimum answer score | 5 upvotes |
123
+ | Accepted answers only | Yes |
124
+ | Minimum answer length | 30 words |
125
+ | Excluded question types | lifestyle, resource lists, book recommendations, tooling |
126
+
127
+ ## Data Collection
128
+
129
+ - **Source**: Stack Exchange API + data dump (archive.org, 2024-12-31 release)
130
+ - **License**: CC-BY-SA 4.0 — free to use with attribution
131
+ - **Distractors**: Real expert-written answers from other questions in the same domain
132
+ - **No synthetic generation**: every word in every answer comes from a real human expert
133
+
134
+ ## Citation
135
+
136
+ ```bibtex
137
+ @dataset{probench2025,
138
+ title = {ProBench: Expert-Level MCQ Benchmark across 5 Professional Domains},
139
+ year = {2025},
140
+ note = {Sourced from Stack Exchange communities (CC-BY-SA 4.0)},
141
+ url = {https://huggingface.co/datasets/lin99/ProBench}
142
+ }
143
+
144
+ @misc{stackexchange_dump2024,
145
+ title = {Stack Exchange Data Dump},
146
+ author = {{Stack Exchange, Inc.}},
147
+ year = {2024},
148
+ url = {https://archive.org/details/stackexchange},
149
+ note = {Licensed under CC-BY-SA 4.0}
150
+ }
151
+ ```
152
+
153
+ ## License
154
+
155
+ All content is from Stack Exchange and licensed under **CC-BY-SA 4.0**.
156
+ Attribution to Stack Exchange communities is required. Each record includes the original URL.