Data

Browse the data

Every (task, model) outcome in VibeSec v1. Each of the 1,000 tasks is a stable URL with the prompt, the generated app, the exploit that proves the bug, and the reference patch. Pivot the heatmap, scan the catalog, or open the trial table.

v1latest

Heatmap Tasks Trials Dataset

Load it in Python

from datasets import load_dataset

ds = load_dataset("muence/vibesec")

View on Hugging Face

Raw files

The exact artifacts this site is built from. Newline-delimited JSON.

tasks.jsonl

One row per task: prompt, app files, spec test, exploit, exploit output, reference patch.

12.7 MB

results.jsonl

One row per (task, model) trial: outcome and failure stage.

717 KB

meta.json

Counts, model list, class distribution and provenance for this release.

1 KB

What a task contains

seed_prompt: The casual, founder-style ask the app was generated from.
app_files: The generated app. It works, and it is vulnerable.
spec_test: Golden-path test. A patch has to keep this passing.
exploit: Runnable attack. Prints PWNED while the bug is present.
patched_files: A known-good reference fix.
vuln_class: Vulnerability class label.

A task only enters the set when the app boots and passes its spec, the exploit prints PWNED, the reference patch kills the exploit, and the patched app still passes. No model judges anything.