Hugging Face's logo

Datasets:
lin99
/
ProBench

Tasks:
Question Answering
Text Generation
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
benchmark
professional
expert
reasoning
law
medicine
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ProBench
File size: 4,394 Bytes
239a5e9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-sa-4.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - benchmark
  - professional
  - expert
  - reasoning
  - law
  - medicine
  - finance
  - cybersecurity
  - engineering
  - llm-evaluation
  - multiple-choice
size_categories:
  - 1K<n<10K
---

# ProBench

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%.

## Why ProBench?

| Benchmark | Frontier model score | Status |
|---|---|---|
| MMLU | >90% | Saturated |
| HellaSwag | >95% | Saturated |
| HumanEval | >90% | Saturated |
| **ProBench** | TBD | Active |

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.

## Dataset Structure

### Splits

| Split | Count |
|---|---|
| train | 370 |
| validation | 79 |
| test | 80 |
| **Total** | **529** |

### By Domain

| Domain | Source Community | Items |
|---|---|---|
| Cybersecurity | security.stackexchange.com | 148 |
| Law | law.stackexchange.com | 121 |
| Medicine | medicalsciences.stackexchange.com | 63 |
| Engineering | engineering.stackexchange.com | 113 |
| Finance | quant.stackexchange.com | 84 |

### Record Format

```json
{
  "id": "cybersecurity_29988",
  "domain": "cybersecurity",
  "question_title": "What is certificate pinning?",
  "question_body": "I've recently seen ...",
  "question_score": 373,
  "question_tags": ["tls", "certificates", "public-key-infrastructure"],
  "choices": {
    "A": "You can roll your own, but you probably will make a major security...",
    "B": "Software is too complex. This is by far the most important factor...",
    "C": "The known_hosts file lets the client authenticate the server...",
    "D": "Typically certificates are validated by checking the signature hierarchy..."
  },
  "answer": "D",
  "distractor_source": "same_domain_answer_pool",
  "source": "stackexchange",
  "license": "CC-BY-SA 4.0",
  "url": "https://security.stackexchange.com/questions/29988/..."
}
```

## Evaluation

Evaluation is exact-match (0 or 1 per question) — fully automated, no human judge required.

```python
from datasets import load_dataset

ds = load_dataset("lin99/ProBench", split="test")

def evaluate(model, dataset):
    correct = 0
    for row in dataset:
        prompt = f"""Domain: {row['domain']}

Question: {row['question_title']}

{row['question_body']}

A. {row['choices']['A']}
B. {row['choices']['B']}
C. {row['choices']['C']}
D. {row['choices']['D']}

Answer (A/B/C/D):"""
        prediction = model(prompt).strip()[0].upper()
        if prediction == row["answer"]:
            correct += 1
    return correct / len(dataset)

accuracy = evaluate(your_model, ds)
print(f"ProBench accuracy: {accuracy:.1%}")
```

## Quality Filters

| Filter | Threshold |
|---|---|
| Minimum question score | 10 upvotes |
| Minimum answer score | 5 upvotes |
| Accepted answers only | Yes |
| Minimum answer length | 30 words |
| Excluded question types | lifestyle, resource lists, book recommendations, tooling |

## Data Collection

- **Source**: Stack Exchange API + data dump (archive.org, 2024-12-31 release)
- **License**: CC-BY-SA 4.0 — free to use with attribution
- **Distractors**: Real expert-written answers from other questions in the same domain
- **No synthetic generation**: every word in every answer comes from a real human expert

## Citation

```bibtex
@dataset{probench2025,
  title     = {ProBench: Expert-Level MCQ Benchmark across 5 Professional Domains},
  year      = {2025},
  note      = {Sourced from Stack Exchange communities (CC-BY-SA 4.0)},
  url       = {https://huggingface.co/datasets/lin99/ProBench}
}

@misc{stackexchange_dump2024,
  title  = {Stack Exchange Data Dump},
  author = {{Stack Exchange, Inc.}},
  year   = {2024},
  url    = {https://archive.org/details/stackexchange},
  note   = {Licensed under CC-BY-SA 4.0}
}
```

## License

All content is from Stack Exchange and licensed under **CC-BY-SA 4.0**.
Attribution to Stack Exchange communities is required. Each record includes the original URL.