JackertheHacker/svg-stack-qwen-captioned-subset

---
configs:
- config_name: default
data_files:
- split: train
path: data/train-*.parquet
- split: validation
path: data/validation-*.parquet
- split: test
path: data/test-*.parquet
---

# SVG Stack Qwen Captioned Subset

This dataset is a relabeled subset derived from StarVector's
`starvector/svg-stack` dataset. It pairs SVG source code with generated natural
language captions describing the rendered visual appearance of each SVG.

The dataset is intended for supervised fine-tuning of text-to-SVG generation
models. A typical training format is to use `Caption` as the input prompt and
`Svg` as the target completion.

Each row contains:

- `Filename`: unique SVG filename from the source dataset
- `Svg`: SVG source code
- `Caption`: generated visual caption

Rows with missing, skipped, or empty captions have been excluded.

## Splits

| Split | Rows | Files |
|---|---:|---|
| train | 340,394 | `data/train-00000-of-00002.parquet`, `data/train-00001-of-00002.parquet` |
| validation | 51,063 | `data/validation-00000-of-00001.parquet` |
| test | 5,664 | `data/test-00000-of-00001.parquet` |

## Captioning

Captions were generated from 384px JPEG renders of the SVGs using
`unsloth/Qwen3.6-35B-A3B`.

Generation settings:

- max tokens: 125
- temperature: 0.7
- top p: 0.8
- top k: 20
- presence penalty: 1.5
- reasoning: off

## Intended Use

Use this dataset to train or evaluate models that generate SVG code from a
short visual description.