VCL VibeBench
Collection
Practical AI model comparisons you can inspect and re-run. Fun, Dev, and Score prompts plus maintainer results — not a blended leaderboard. Stop trust • 2 items • Updated
Datasets:
id stringclasses 10
values | suite stringclasses 1
value | suite_version stringclasses 1
value | title stringclasses 10
values | prompt stringclasses 10
values | scoring stringclasses 2
values | expected_hint stringclasses 5
values |
|---|---|---|---|---|---|---|
dev-001-find-a-subtle-bug-off-by-one | dev | 1.1 | Find a subtle bug (off-by-one) | This function is supposed to return the last 3 items of a list, but it has a bug.
Find the bug and explain the fix in 2 sentences, then show the corrected code.
def last_three(items):
return items[-4:-1] | scored | -3 |
dev-002-read-a-real-stack-trace | dev | 1.1 | Read a real stack trace | A user reports their Node.js app crashes on startup with this error. What's the most likely
cause and the first thing you'd check?
TypeError: Cannot read properties of undefined (reading 'map')
at renderList (/app/src/components/List.jsx:14:23)
at processTicksAndRejections (node:internal/process/task_queues:95... | exploratory | null |
dev-003-refactor-for-readability | dev | 1.1 | Refactor for readability | Refactor this JavaScript into cleaner, more idiomatic code without changing its behavior.
Briefly explain what you changed and why.
function f(a,b,c){if(a==true){if(b>c){return b}else{return c}}else{return a}} | exploratory | null |
dev-004-write-tests-for-edge-cases | dev | 1.1 | Write tests for edge cases | Write 3 unit test cases (in plain pseudocode) for a function `divide(a, b)` that should
catch the trickiest bugs a beginner would miss. | exploratory | null |
dev-005-fix-a-broken-sql-query | dev | 1.1 | Fix a broken SQL query | This query is supposed to return each customer's total spend, but it's returning wrong
totals for customers with multiple orders. What's wrong and what's the fix?
SELECT customers.name, orders.amount
FROM customers
JOIN orders ON customers.id = orders.customer_id; | scored | GROUP BY |
dev-006-multi-step-instruction-following | dev | 1.1 | Multi-step instruction following | I need a Python function called `slugify` that: (1) lowercases the input, (2) replaces
spaces with hyphens, (3) strips any character that isn't a-z, 0-9, or a hyphen, and
(4) collapses multiple hyphens into one. Give me the function and one example call. | scored | def slugify |
dev-007-explain-a-tradeoff-architecture-judgment | dev | 1.1 | Explain a tradeoff (architecture judgment) | I'm building a side project SaaS with an indie budget. Should I use a queue-based
background job system or just cron a script every few minutes for sending scheduled
emails to ~500 users/day? Give me a direct recommendation, not a list of options. | exploratory | null |
dev-008-spot-the-security-bug-xss | dev | 1.1 | Spot the security bug (XSS) | This Express route renders a user-supplied name into HTML. Name the vulnerability in one
word and the safest fix in one sentence, then show a corrected one-liner.
app.get('/hello', (req, res) => {
res.send('<h1>Hello ' + req.query.name + '</h1>');
}); | scored | xss |
dev-009-email-validation-with-regex | dev | 1.1 | Email validation with regex | Write a JavaScript function `isValidEmail(email)` that returns true only for a basic
well-formed email address using a regular expression. Keep it short; include the function. | scored | function |
dev-010-find-the-async-race | dev | 1.1 | Find the async race | This Node.js snippet is supposed to print 1000, but often prints less. What is the bug,
and how do you fix it? Explain in 3 sentences or less, then show corrected code.
let count = 0;
async function bump() {
const next = count + 1;
await Promise.resolve();
count = next;
}
await Promise.all(Array.from({ length: 1... | exploratory | null |
VCL VibeBench
Stop trusting benchmark slides. Run it yourself.
Practical AI model prompts from Vibe Coder's Life.
This dataset mirrors the open-source GitHub suites. It is not a blended intelligence leaderboard. There is no LLM judge. 9/12 on Score means nine JavaScript helpers compiled and passed hidden unit tests — not “75% smart.”
Configs
| Config | What it is | Rows |
|---|---|---|
fun |
Fun 1.0 — short copy-paste prompts | 10 |
dev |
Dev 1.1 — coding / debugging prompts | 10 |
score |
Score 1.2 — write one named JS function | 12 |
results |
Slim maintainer run summaries (pass/fail counts) | varies |
from datasets import load_dataset
fun = load_dataset("kondasviktor/vcl-vibebench", "fun")
dev = load_dataset("kondasviktor/vcl-vibebench", "dev")
score = load_dataset("kondasviktor/vcl-vibebench", "score")
results = load_dataset("kondasviktor/vcl-vibebench", "results")
Hidden tests stay on GitHub
Score prompts are here. Hidden unit tests live in the GitHub harness. Fork and run:
git clone https://github.com/kondasviktor/vcl-ai-model-arena.git
cd vcl-ai-model-arena
cp .env.example .env # OPENROUTER_API_KEY
npm ci
MODELS=google/gemini-3.6-flash npm run eval:score:smoke
Links
- Interactive explorer (Static Space): kondasviktor/vcl-vibebench
- Source + CLI: github.com/kondasviktor/vcl-ai-model-arena
- Write-ups + newsletter: vibecoderslife.com/vibebench
Prompts and result summaries: CC BY 4.0. Code on GitHub: MIT.
- Downloads last month
- 12