codemetic commited on
Commit
707db2c
·
verified ·
1 Parent(s): 9928543

Remove dataset card (data files only)

Browse files
Files changed (1) hide show
  1. README.md +0 -120
README.md DELETED
@@ -1,120 +0,0 @@
1
- ---
2
- configs:
3
- - config_name: megavul
4
- default: true
5
- task_categories:
6
- - text-classification
7
- tags:
8
- - code
9
- - security
10
- - vulnerability-detection
11
- - vulnerable-fixed-pairs
12
- - c
13
- - cpp
14
- - java
15
- - multilingual
16
- pretty_name: LAMBDA
17
- size_categories:
18
- - 10K<n<100K
19
- ---
20
- # LAMBDA -- Vulnerability detection benchmarks with vulnerable-fixed pairs
21
-
22
- Benchmark data of the **LAMBDA** framework (LLM-Augmented Methodology for
23
- Vulnerability Detection via Formal Assertions). Three subsets, each split
24
- 8:1:1 into `train` / `val` / `test` (project- and temporally-separated):
25
-
26
- | config | train | val | test | languages |
27
- | --- | --- | --- | --- | --- |
28
- | `primevul` | 3731 | 467 | 466 | C/C++ only |
29
- | `megavul` | 4304 | 538 | 538 | train = C/C++, val/test = Java |
30
- | `crossvul` | 4413 | 552 | 551 | train = {C, Go, Java, PHP, TypeScript}; val/test = {C++, C#, JavaScript, Python, Ruby, Rust, Scala} |
31
-
32
- Each row is a **function-level vulnerable-fixed pair**: the vulnerable code
33
- before a security fix and the fixed code after it, plus metadata (CWE/CVE,
34
- project, commit, language). `train` is used to induce the LAMBDA knowledge
35
- base; `val` for hyperparameter selection; `test` for final evaluation.
36
-
37
- ## Subsets
38
-
39
- ### `primevul` (single-language, C/C++)
40
-
41
- | field | description |
42
- | --- | --- |
43
- | idx | row index |
44
- | project | source project |
45
- | commit_id | fix commit |
46
- | project_url | project URL |
47
- | commit_url | commit URL |
48
- | commit_message | commit message |
49
- | target | label (1 = vulnerable) |
50
- | vul_code | vulnerable function (before the fix) |
51
- | fixed_code | fixed function (after the fix) |
52
- | func_hash | function hash |
53
- | file_name | source file name |
54
- | file_hash | file hash |
55
- | cwe | CWE ids |
56
- | cve | CVE id |
57
- | cve_desc | CVE description |
58
- | nvd_url | NVD URL |
59
-
60
- Label column: `target` (1 = vulnerable). All code is C/C++.
61
-
62
- ### `megavul` (cross-language: C/C++ -> Java)
63
-
64
- | field | description |
65
- | --- | --- |
66
- | cve_id | CVE id |
67
- | cwe | CWE id |
68
- | publish_date | CVE publish date |
69
- | repo_name | source repository |
70
- | commit_msg | fix commit message |
71
- | vul_code | vulnerable function (before the fix) |
72
- | abstract_func_before | abstracted vulnerable function |
73
- | fixed_code | fixed function (after the fix) |
74
- | abstract_func | abstracted fixed function |
75
- | abstract_symbol_table | abstract symbol table |
76
- | func_graph_path | function graph path |
77
- | is_vul | label (true = vulnerable) |
78
-
79
- Label column: `is_vul`. `train` contains C/C++ samples, `val`/`test` Java
80
- samples (cross-language transfer evaluation).
81
-
82
- ### `crossvul` (7 unseen languages in val/test)
83
-
84
- | field | description |
85
- | --- | --- |
86
- | cwe_id | CWE id |
87
- | cwe_description | CWE description |
88
- | language | programming language |
89
- | vul_code | vulnerable code (before fix) |
90
- | fixed_code | fixed code (after fix) |
91
- | file_pair_id | file-pair id |
92
- | source | source dataset |
93
- | language_dir | language directory |
94
-
95
- No label column (all rows are vulnerable-fixed pairs); train languages are
96
- {C, Go, Java, PHP, TypeScript}, val/test languages are
97
- {C++, C#, JavaScript, Python, Ruby, Rust, Scala} -- no overlap between the
98
- training and held-out language sets.
99
-
100
- ## Load
101
-
102
- ```python
103
- from datasets import load_dataset
104
- ds = load_dataset("codemetic/LAMBDA", "primevul", split="test")
105
- ```
106
-
107
- ## Provenance
108
-
109
- This repository aggregates three public vulnerability datasets (rows are
110
- redistributed unchanged, only the 8:1:1 split is applied):
111
-
112
- - **primevul** -- PrimeVul: "PrimeVul: Towards an Era of AI-Assisted Software Vulnerability Detection" (Ding et al., 2025), https://huggingface.co/datasets/PublicState/PrimeVul
113
- - **megavul** -- MegaVul: "Can Large Language Models Find Software Vulnerabilities?" (Ni et al., 2024), https://github.com/Icyrockton/MegaVul
114
- - **crossvul** -- CrossVul: "A Cross-Language Vulnerability Dataset with Severity Data" (Nikitopoulos et al., 2021), https://github.com/fabsx00/crossvul
115
-
116
- ## License
117
-
118
- The constituent datasets retain their original licenses; this repository
119
- imposes no additional license. Please cite the source datasets listed above
120
- when using this data.