README.md

---
dataset_info:
  features:
  - name: Filename
    dtype: string
  - name: Svg
    dtype: string
  - name: image
    dtype: image
  - name: width
    dtype: float32
  - name: height
    dtype: float32
  - name: svgo
    struct:
    - name: original_size
      dtype: int32
    - name: optimized_size
      dtype: int32
    - name: bytes_saved
      dtype: int32
    - name: saved_ratio
      dtype: float32
    - name: svgo_error
      dtype: string
  - name: svg_token_count
    dtype: int64
  - name: prompt
    dtype: string
  - name: completion
    dtype: string
  splits:
  - name: train
    num_bytes: 30372678141.375
    num_examples: 1440517
  - name: test
    num_bytes: 1621372928.375
    num_examples: 76629
  - name: val
    num_bytes: 1620844535.5
    num_examples: 76652
  download_size: 31773499267
  dataset_size: 33614895605.25
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: val
    path: data/val-*
---

# Dataset Card for svg-stack-filtered

This is an attempt to replicate the dataset used for SFT in the paper
[Rendering-Aware Reinforcement Learning for Vector Graphics Generation](https://huggingface.co/papers/2505.20793)

Processed:
- Optimized with svgo precision=2
- Rasterized with cairosvg[^cairo]

[^cairo] cairosvg doesn't implement all svg features, but matches how the original paper

Filtered based on some heuristics:
- Removed any svg that couldn't be rendered with cairosvg (~30%)
- Removed solid-color images
- Removed some problematic tags
- Removed svgs with `<text>`
- Removed svgs with embedded data url images
- Removed any svgs under 100 tokens
- Removed some broken svgs with nested `<svg>` tags