Datasets:

Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
cogwheelhead commited on
Commit
15423f2
·
verified ·
1 Parent(s): 9504ba6

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +96 -3
README.md CHANGED
@@ -1,3 +1,96 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ pretty_name: μ-MATH
4
+ task_categories:
5
+ - text-generation
6
+ language:
7
+ - en
8
+ tags:
9
+ - math
10
+ - reasoning
11
+ size_categories:
12
+ - 1K<n<10K
13
+ dataset_info:
14
+ features:
15
+ - name: uuid
16
+ dtype: string
17
+ - name: subject
18
+ dtype: string
19
+ - name: has_image
20
+ dtype: bool
21
+ - name: image
22
+ dtype: binary
23
+ - name: problem_statement
24
+ dtype: string
25
+ - name: golden_answer
26
+ dtype: string
27
+ splits:
28
+ - name: test
29
+ num_examples: 1100
30
+ configs:
31
+ - config_name: default
32
+ data_files:
33
+ - split: test
34
+ path: data/test-*
35
+ ---
36
+
37
+ **U-MATH** is a comprehensive benchmark of 1,100 unpublished university-level problems sourced from real teaching materials.
38
+
39
+ It is designed to evaluate the mathematical reasoning capabilities of Large Language Models (LLMs). \
40
+ The dataset is balanced across six core mathematical topics and includes 20% of multimodal problems (involving visual elements such as graphs and diagrams).
41
+
42
+ For fine-grained performance evaluation results and detailed discussion, check out our [paper](LINK).
43
+
44
+ * 📊 [U-MATH benchmark at Hugginface](https://huggingface.co/datasets/toloka/umath)
45
+ * 🔎 [μ-MATH benchmark at Hugginface](https://huggingface.co/datasets/toloka/mumath)
46
+ * 🗞️ [Paper](LINK)
47
+ * 👾 [Evaluation Code at GitHub](https://github.com/Toloka/u-math/)
48
+
49
+ ### Key Features
50
+ * **Topics Covered**: Precalculus, Algebra, Differential Calculus, Integral Calculus, Multivariable Calculus, Sequences & Series.
51
+ * **Problem Format**: Free-form answer with LLM judgement
52
+ * **Evaluation Metrics**: Accuracy; splits by subject and text-only vs multimodal problem type.
53
+ * **Curation**: Original problems composed by math professors and used in university curricula, samples validated by math experts at [Toloka AI](https://toloka.ai), [Gradarius](https://www.gradarius.com)
54
+
55
+ ### Dataset Fields
56
+ `uuid`: problem id \
57
+ `has_image`: a boolean flag on whether the problem is multimodal or not \
58
+ `image`: binary data encoding the accompanying image, empty for text-only problems \
59
+ `subject`: subject tag marking the topic that the problem belongs to \
60
+ `problem_statement`: problem formulation, written in natural language \
61
+ `golden_answer`: a correct solution for the problem, written in natural language \
62
+
63
+ For meta-evaluation (evaluating the quality of LLM judges), refer to [µ-MATH dataset](https://huggingface.co/datasets/toloka/mu-math).
64
+
65
+ ### Evaluation Results
66
+
67
+ <div align="center">
68
+ <img src="https://cdn-uploads.huggingface.co/production/uploads/650238063e61bc019201e3e2/beMyOikpKfp3My2vu5Mjc.png" alt="umath-table" width="800"/>
69
+ </div>
70
+
71
+ <div align="center">
72
+ <img src="https://cdn-uploads.huggingface.co/production/uploads/650238063e61bc019201e3e2/7_VZXidxMHG7PiDM983lS.png" alt="umath-bar" width="950"/>
73
+ </div>
74
+
75
+ The prompt used for inference:
76
+
77
+ ```
78
+ {problem_statement}
79
+ Please reason step by step, and put your final answer within \boxed{}
80
+ ```
81
+
82
+ ### Licensing Information
83
+ All the dataset contents are available under the MIT license.
84
+
85
+ ### Citation
86
+ If you use U-MATH or μ-MATH in your research, please cite the paper:
87
+ ```bibtex
88
+ @inproceedings{umath2024,
89
+ title={U-MATH: A University-Level Benchmark for Evaluating Mathematical Skills in LLMs},
90
+ author={Konstantin Chernyshev, Vitaliy Polshkov, Ekaterina Artemova, Alex Myasnikov, Vlad Stepanov, Alexei Miasnikov and Sergei Tilga},
91
+ year={2024}
92
+ }
93
+ ```
94
+
95
+ ### Contact
96
+ For inquiries, please contact [email protected]