CVE-Backport-Qwen2.5-Coder-32B / eval /test-generation-v4-n50.json
anicka's picture
Add eval results: eval/test-generation-v4-n50.json
d403cc1 verified
{
"module": "domain/cve-backport",
"eval_type": "test-generation",
"timestamp": "2026-03-28T10:06:38Z",
"n_examples": 50,
"metrics": {
"avg_score": 0.6661,
"median_score": 0.802,
"good_tests": 50,
"errors": 0,
"zero_errors": true
},
"pass": true,
"per_example": [
{
"id": "example_0",
"cve": "CVE-2020-27749",
"package": "grub2",
"tier": "identical",
"score": 0.485,
"ref_overlap": 0.321,
"prompt_overlap": 0.122,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 14
},
{
"id": "example_14",
"cve": "CVE-2020-1752",
"package": "glibc",
"tier": "identical",
"score": 0.455,
"ref_overlap": 0.286,
"prompt_overlap": 0.062,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 9
},
{
"id": "example_28",
"cve": "CVE-2019-9674",
"package": "python",
"tier": "identical",
"score": 0.367,
"ref_overlap": 0.091,
"prompt_overlap": 0.106,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 7
},
{
"id": "example_42",
"cve": "CVE-2019-1551",
"package": "openssl-1_1",
"tier": "identical",
"score": 0.75,
"ref_overlap": 0.857,
"prompt_overlap": 0.109,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 11
},
{
"id": "example_56",
"cve": "CVE-2020-10029",
"package": "glibc",
"tier": "identical",
"score": 0.358,
"ref_overlap": 0.111,
"prompt_overlap": 0.01,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_70",
"cve": "CVE-2024-53907",
"package": "python-Django",
"tier": "identical",
"score": 0.822,
"ref_overlap": 1.0,
"prompt_overlap": 0.109,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 21
},
{
"id": "example_84",
"cve": "CVE-2019-3885",
"package": "pacemaker",
"tier": "identical",
"score": 0.328,
"ref_overlap": 0.05,
"prompt_overlap": 0.014,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 5
},
{
"id": "example_98",
"cve": "CVE-2021-3933",
"package": "openexr",
"tier": "identical",
"score": 0.397,
"ref_overlap": 0.188,
"prompt_overlap": 0.014,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_112",
"cve": "CVE-2020-8286",
"package": "curl",
"tier": "identical",
"score": 0.671,
"ref_overlap": 0.727,
"prompt_overlap": 0.039,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 11
},
{
"id": "example_126",
"cve": "CVE-2021-22924",
"package": "curl",
"tier": "identical",
"score": 0.507,
"ref_overlap": 0.4,
"prompt_overlap": 0.037,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 10
},
{
"id": "example_140",
"cve": "CVE-2019-7397",
"package": "GraphicsMagick",
"tier": "identical",
"score": 0.455,
"ref_overlap": 0.3,
"prompt_overlap": 0.025,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_154",
"cve": "CVE-2018-0495",
"package": "libgcrypt",
"tier": "identical",
"score": 0.363,
"ref_overlap": 0.12,
"prompt_overlap": 0.015,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_168",
"cve": "CVE-2017-12133",
"package": "glibc",
"tier": "identical",
"score": 0.712,
"ref_overlap": 0.778,
"prompt_overlap": 0.114,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 10
},
{
"id": "example_182",
"cve": "CVE-2017-6441",
"package": "php7",
"tier": "identical",
"score": 0.554,
"ref_overlap": 0.5,
"prompt_overlap": 0.021,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_196",
"cve": "CVE-2015-8994",
"package": "php7",
"tier": "identical",
"score": 0.804,
"ref_overlap": 1.0,
"prompt_overlap": 0.018,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_210",
"cve": "CVE-2016-8866",
"package": "GraphicsMagick",
"tier": "identical",
"score": 0.472,
"ref_overlap": 0.333,
"prompt_overlap": 0.024,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_224",
"cve": "CVE-2018-19541",
"package": "jasper",
"tier": "synthetic-adapted",
"score": 0.517,
"ref_overlap": 0.429,
"prompt_overlap": 0.014,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_238",
"cve": "CVE-2019-9636",
"package": "python3-base",
"tier": "synthetic-adapted",
"score": 0.806,
"ref_overlap": 1.0,
"prompt_overlap": 0.03,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_252",
"cve": "CVE-2017-15597",
"package": "xen",
"tier": "synthetic-adapted",
"score": 0.368,
"ref_overlap": 0.133,
"prompt_overlap": 0.008,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_266",
"cve": "CVE-2018-0735",
"package": "openssl-1_1",
"tier": "adapted",
"score": 0.805,
"ref_overlap": 1.0,
"prompt_overlap": 0.023,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_280",
"cve": "CVE-2021-27219",
"package": "glib2",
"tier": "identical",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.015,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_294",
"cve": "CVE-2018-5711",
"package": "gd",
"tier": "identical",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.016,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_308",
"cve": "CVE-2019-11041",
"package": "php7",
"tier": "synthetic-adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.014,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_322",
"cve": "CVE-2022-0113",
"package": "chromium",
"tier": "synthetic-adapted",
"score": 0.802,
"ref_overlap": 1.0,
"prompt_overlap": 0.009,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_336",
"cve": "CVE-2019-17498",
"package": "libssh2_org",
"tier": "identical",
"score": 0.804,
"ref_overlap": 1.0,
"prompt_overlap": 0.019,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_350",
"cve": "CVE-2019-5868",
"package": "chromium",
"tier": "identical",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.016,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_364",
"cve": "CVE-2018-17456",
"package": "git",
"tier": "identical",
"score": 0.604,
"ref_overlap": 1.0,
"prompt_overlap": 0.021,
"structural": 0.333,
"has_assert": false,
"has_function": false,
"lines": 3
},
{
"id": "example_378",
"cve": "CVE-2017-3737",
"package": "openssl",
"tier": "identical",
"score": 0.736,
"ref_overlap": 0.818,
"prompt_overlap": 0.136,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 13
},
{
"id": "example_392",
"cve": "CVE-2021-3500",
"package": "djvulibre",
"tier": "synthetic-adapted",
"score": 0.806,
"ref_overlap": 1.0,
"prompt_overlap": 0.029,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_406",
"cve": "CVE-2019-12838",
"package": "slurm",
"tier": "synthetic-adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.016,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_420",
"cve": "CVE-2016-9042",
"package": "ntp",
"tier": "synthetic-adapted",
"score": 0.804,
"ref_overlap": 1.0,
"prompt_overlap": 0.019,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_434",
"cve": "CVE-2018-5732",
"package": "dhcp",
"tier": "synthetic-adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.016,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_448",
"cve": "CVE-2019-2614",
"package": "mariadb",
"tier": "identical",
"score": 0.606,
"ref_overlap": 1.0,
"prompt_overlap": 0.029,
"structural": 0.333,
"has_assert": false,
"has_function": false,
"lines": 4
},
{
"id": "example_462",
"cve": "CVE-2019-9947",
"package": "python-base",
"tier": "identical",
"score": 0.805,
"ref_overlap": 1.0,
"prompt_overlap": 0.026,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_476",
"cve": "CVE-2018-9135",
"package": "ImageMagick",
"tier": "adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.017,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_490",
"cve": "CVE-2017-9798",
"package": "apache2",
"tier": "adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.013,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_504",
"cve": "CVE-2016-9317",
"package": "gd",
"tier": "identical",
"score": 0.805,
"ref_overlap": 1.0,
"prompt_overlap": 0.027,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_518",
"cve": "CVE-2018-20852",
"package": "python-doc",
"tier": "synthetic-adapted",
"score": 0.806,
"ref_overlap": 1.0,
"prompt_overlap": 0.031,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_532",
"cve": "CVE-2016-9603",
"package": "xen",
"tier": "synthetic-adapted",
"score": 0.802,
"ref_overlap": 1.0,
"prompt_overlap": 0.008,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_546",
"cve": "CVE-2020-9490",
"package": "apache2",
"tier": "synthetic-adapted",
"score": 0.802,
"ref_overlap": 1.0,
"prompt_overlap": 0.011,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_560",
"cve": "CVE-2017-3144",
"package": "dhcp",
"tier": "synthetic-adapted",
"score": 0.802,
"ref_overlap": 1.0,
"prompt_overlap": 0.012,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_574",
"cve": "CVE-2017-17479",
"package": "openjpeg2",
"tier": "synthetic-adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.017,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_588",
"cve": "CVE-2021-3781",
"package": "ghostscript-mini",
"tier": "identical",
"score": 0.804,
"ref_overlap": 1.0,
"prompt_overlap": 0.018,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_602",
"cve": "CVE-2018-20103",
"package": "haproxy",
"tier": "identical",
"score": 0.805,
"ref_overlap": 1.0,
"prompt_overlap": 0.026,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_616",
"cve": "CVE-2016-8862",
"package": "ImageMagick",
"tier": "identical",
"score": 0.804,
"ref_overlap": 1.0,
"prompt_overlap": 0.022,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_630",
"cve": "CVE-2020-13977",
"package": "nagios",
"tier": "identical",
"score": 0.809,
"ref_overlap": 1.0,
"prompt_overlap": 0.043,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 3
},
{
"id": "example_644",
"cve": "CVE-2018-1060",
"package": "python3",
"tier": "synthetic-adapted",
"score": 0.803,
"ref_overlap": 1.0,
"prompt_overlap": 0.017,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
},
{
"id": "example_658",
"cve": "CVE-2018-10931",
"package": "cobbler",
"tier": "synthetic-adapted",
"score": 0.446,
"ref_overlap": 0.25,
"prompt_overlap": 0.105,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 10
},
{
"id": "codegen-curl-10079-libcurl-ocloexec.patch-cookie.c",
"cve": "CVE-2018-16839",
"package": "curl",
"tier": "identical",
"score": 0.322,
"ref_overlap": 0.0,
"prompt_overlap": 0.111,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 9
},
{
"id": "codegen-python3-11362-CVE-2019-16056-email-parse-addr.patch-Lib_email__header_value_parser.py",
"cve": "CVE-2019-16056",
"package": "python3",
"tier": "identical",
"score": 0.307,
"ref_overlap": 0.0,
"prompt_overlap": 0.036,
"structural": 1.0,
"has_assert": true,
"has_function": true,
"lines": 4
}
]
}