Hugging Face's logo

Datasets:
FuCY
/
php

Tasks:
Text Generation
Modalities:
Text
Formats:
text
Languages:
English
Size:
1K - 10K
Tags:
cybersecurity
vulnerability
php
benchmark
cybergym
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
php / README.md
FuCY's picture
FuCY
Upload folder using huggingface_hub
1316a2e verified
preview code
|
Raw History Blame Contribute Delete
4.05 kB
metadata 
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - cybersecurity
  - vulnerability
  - php
  - benchmark
  - cybergym

Cybergym-Pro — PHP benchmark tasks

English | 中文

Snapshot: 2026-11-04. This dataset contains the PHP portion of the Cybergym-Pro benchmark: 199 task packages, each shipped with the same four-level material set used by the other language datasets in the same collection.

Layout 

tasks/<task_id>/
  images/<task_id>-vul.tar.gz     # docker image archive, vulnerable build
  images/<task_id>-fix.tar.gz     # docker image archive, fixed build
  levels/level0/                  # repo-vul.tar.gz + README.md + submit.sh
  levels/level1/                  # + description.txt
  levels/level2/                  # + error.txt (reference input output on vul)
  levels/level3/                  # + patch.diff and repo-fix.tar.gz
  source/repo-vul.tar.gz          # vulnerable source snapshot
  source/repo-fix.tar.gz          # fixed source snapshot
  source/patch.diff               # upstream fix diff
  source/source-hashes.json       # archive hashes
  metadata.json                   # task id, CVE ids, upstream repo/commits, image tags
agent/cybergym-agent-php-1.tar.gz # PHP agent runtime image (docker save, gzipped)

<task_id> is the 16-hex Cybergym task id (e.g. 019e0f0b88563440); the CVE mapping is in each task's metadata.json (cve_ids).

Level semantics

  • level0: vulnerable source + README.md + submit.sh (run a PoC against the vulnerable build).
  • level1: adds description.txt — component and observable behaviour only, no CVE id, fix commit, patch or decision procedure.
  • level2: adds error.txt, the real output of the reference input on the vulnerable program.
  • level3: adds patch.diff and the fixed source archive.

All 199 task packages carry level0–level3; the six delivered PHP L1 batches (10/20/30/40/49/50) are represented exactly once.

Using the images 

docker load -i tasks/<task_id>/images/<task_id>-vul.tar.gz
docker load -i tasks/<task_id>/images/<task_id>-fix.tar.gz
docker run --rm --network none -v /path/to/poc:/tmp/poc:ro \
  cybergym/php:<task_id>-vul /usr/local/bin/run_poc /tmp/poc

Each task exposes the same runtime contract: /usr/local/bin/run_poc <poc> executes the submitted PoC inside the task image. Detection is per-task and is derived from real vulnerable/fixed behaviour, never from exit status alone.

Agent runtime

agent/cybergym-agent-php-1.tar.gz is the gzipped docker save archive of the PHP agent image (cybergym-agent/php:1). It bundles PHP 8.4, Composer, Python, Node.js, Codex CLI, Claude Code and OpenHands.

gunzip -c agent/cybergym-agent-php-1.tar.gz | docker load
docker run --rm -it cybergym-agent/php:1

Notes

  • Reference PoCs, evaluation contracts and admin validation material are kept private and are not part of this dataset.
  • metadata.json records provenance only (upstream repository, vulnerable/fix commits, image tags, reference PoC hash); it contains no PoC source.
  • Task packages are provided for benchmark evaluation. Use them only in sandboxed environments you control.

中文说明

本数据集是 Cybergym-Pro 的 PHP 部分，共 199 个任务包，结构与同一 collection 中的其他语言数据集一致：

  • tasks/<task_id>/：任务包，含漏洞/修复两份镜像归档（images/）、四层材料 （levels/level0–level3）、源码快照与补丁（source/）、元数据（metadata.json）；
  • agent/cybergym-agent-php-1.tar.gz：PHP agent 运行镜像（docker save 后 gzip）。

level0 只有漏洞源码与提交脚本；level1 增加描述；level2 增加参考输入在漏洞 版本上的真实输出；level3 增加补丁与修复源码。199 个任务全部具备四层材料。

参考 PoC、评测契约与管理端校验材料不在本数据集中，仅保留任务本身所需的源码、 镜像与分层材料。