The dataset viewer is not available for this dataset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
CyberGym task ID splits
Task ID splits used in our work on CyberGym vulnerability-reproduction benchmarks. Each row is a single task identifier (no inputs / no outputs); this dataset is intended as a task-list manifest for downstream evaluation scripts that fetch the actual task workspaces from the CyberGym distribution.
Configs
| Config | Split | Rows | Description |
|---|---|---|---|
full |
train |
1507 | Every task in the CyberGym Level-1 release |
train |
train |
300 | Training pool used in our experiments |
eval |
test |
200 | Held-out evaluation pool |
train and eval are deterministic (project-stratified) subsets of full.
Schema
{
"task_id": "arvo:11504", # canonical CyberGym ID
"kind": "arvo", # "arvo" or "oss-fuzz"
"id": 11504, # numeric portion
"project": "libxml2", # project name from CyberGym's db.json
}
Usage
from datasets import load_dataset
eval_ids = load_dataset("Elfsong/cybergym-tasks", "eval")["test"]["task_id"]
train_ids = load_dataset("Elfsong/cybergym-tasks", "train")["train"]["task_id"]
full_ids = load_dataset("Elfsong/cybergym-tasks", "full")["train"]["task_id"]
Known issue: 8-task TRAIN/EVAL overlap
The train and eval configs share 8 task IDs, all oss-fuzz:*:
| task_id | project |
|---|---|
| oss-fuzz:42536536 | qpdf |
| oss-fuzz:42537493 | libxml2 |
| oss-fuzz:42537664 | ffmpeg |
| oss-fuzz:42537686 | ffmpeg |
| oss-fuzz:42537734 | ffmpeg |
| oss-fuzz:42538131 | ffmpeg |
| oss-fuzz:383170474 | libdwarf |
| oss-fuzz:383825645 | ffmpeg |
Looks like a sampling bug (no de-duplication when drawing the two splits). Downstream evaluators should subtract these from one side if a clean held-out is required.
Provenance & citation
Underlying tasks are drawn from CyberGym (1,507 real-world vulnerability reproduction tasks from ARVO and OSS-Fuzz, 188 projects). Cite the original:
@misc{wang2026cybergym,
title = {{CyberGym}: A Benchmark for LLM-driven Vulnerability Reproduction},
author = {Wang, [...]},
year = {2026}
}
This split list itself is released under CC-BY-4.0; the underlying CyberGym task workspaces follow CyberGym's own license.
- Downloads last month
- 351