---
license: mit
task_categories:
- text-generation
tags:
- catbench
- evaluation
- exl3
pretty_name: CatBench Results
---

# CatBench results

Model outputs for [CatBench](https://katehuuh.github.io/demos/CatBench/), a small
benchmark that asks a model to draw a cute kitten two ways and looks at what comes
back. Produced by the `/catbench` command in
[blockquant](https://github.com/Honkware/blockquant).

Upstream publishes its own results at
[Katehuuh.github.io/demos/CatBench/assets](https://github.com/Katehuuh/Katehuuh.github.io/tree/main/demos/CatBench/assets).
This dataset holds runs for models that are not in that set. `/catbench` checks both
and only rents a pod when neither has the model, so the two do not duplicate
each other.

## The prompts

Verbatim, unchanged between runs:

1. `Create a detailed SVG image of a cute kitten.`
2. `Write a Python script that draws a cute kitten using matplotlib.`

## Layout

One entry per benched model, keyed by the repo stem lowercased with `_` and
whitespace folded to `-`, matching upstream's key format.

```
manifest.json                 index of every entry
assets/<key>-svg.jpg          the SVG, rasterized
assets/<key>-python.jpg       what the matplotlib script drew
assets/<key>.svg              raw SVG the model emitted
assets/<key>.py               raw script the model emitted
```

Each manifest entry records the model id, both image paths, both source paths,
which loader ran (`exl3` or `transformers`), the engine version, and the run date.

## What is here and what is not

Only runs where both halves produced something. A run whose script crashed, whose
SVG came back empty, or where the model wrote prose instead of code is not stored,
so those can be retried after a fix rather than being cached as a result.

Raw sources are kept alongside the images because the rendering is lossy: a script
that draws a kitten badly and one that draws nothing look more alike as a jpg than
they do as code.

## Caveats

Greedy decoding, one sample per prompt, no retries. This measures what a model does
on one attempt at two prompts. It is a curiosity benchmark, not a capability
evaluation, and it should not be read as one.