DanielRegaladoCardoso/svg-chart-render-v1

SVG Chart Render Mix v1

Training data for fine-tuning a small code model (DeepSeek Coder 1.3B) to map (chart specification JSON) to inline SVG code.

Part of the SQL Agent LLMOps project.

Total	Sources	Input	Output
~25,000 rows	2	structured JSON chart spec	rendered SVG string

Part of the SQL Agent LLMOps project

Dataset	Model	Role
DanielRegaladoCardoso/text-to-sql-mix-v2	Qwen 2.5 Coder 7B	NL question to SQL
DanielRegaladoCardoso/chart-reasoning-mix-v1	Phi-3 Mini 3.8B	(question + result) to chart spec
DanielRegaladoCardoso/svg-chart-render-v1	DeepSeek Coder 1.3B	chart spec to SVG

Schema

Field	Type	Description
id	string	Stable hash-based identifier
chart_spec	string (JSON)	Input: chart type, data points, encoding, title, axis labels
svg_code	string	Target output: full <svg>...</svg> inline SVG
source	string	Origin tag (see source attribution)
metadata	string (JSON)	chart_type, num_points, svg_size_bytes

Note: chart_spec and metadata are stored as JSON strings in the parquet. Parse with json.loads(row["chart_spec"]) after loading.

chart_spec structure

{
  "chart_type": "bar|line|scatter|donut|histogram|area",
  "data": [{"x": "value", "y": "value", "color": "value|null"}],
  "encoding": {"x": "col_name", "y": "col_name", "color": "col_name|null"},
  "title": "string or null",
  "x_label": "string or null",
  "y_label": "string or null"
}

Splits

Split	Rows
train	~23,750
validation	~625
test	~625

Source attribution

Source	Tag in source	Rows	License	Link	Method	Notes
nvBench chart renders	synth-matplotlib	~10,500	Apache-2.0 (pipeline) + MIT (nvBench data)	nvBench GitHub	Programmatic	Each of 7,247 nvBench chart configs rendered via matplotlib's SVG backend. Up to 3 title-augmented variants per entry. Chart types: bar, line, scatter, donut. Perfect (spec, svg) pairs.
svgen-500k filtered	svgen500k-*	~15,000	Per upstream row	umuthopeyildirim/svgen-500k	Filtered	Streamed 216k SVGs, heuristic-filtered to chart-shaped structures (multi-element: has <rect>, <line>, <g>). Rejects single-path icons. Provides general SVG syntax fluency.

Pipeline

Build script: training/data_pipelines/build_svg_mix.py

Three stages:

1. synth-charts -- load nvBench chart configs, replay each in matplotlib (Agg/SVG backend), augment with NL paraphrases as candidate titles.
2. svgen -- stream umuthopeyildirim/svgen-500k, keep only chart-shaped SVGs.
3. combine-push -- dedup by SVG hash, 95/2.5/2.5 split, push to HF with card.

Usage

from datasets import load_dataset
import json

ds = load_dataset("DanielRegaladoCardoso/svg-chart-render-v1")
ex = ds["train"][0]

spec = json.loads(ex["chart_spec"])
print(spec["chart_type"])
print(ex["svg_code"][:200])

SFT format

import json

def to_sft(row):
    return {
        "messages": [
            {"role": "system", "content": "You render chart specifications as inline SVG."},
            {"role": "user", "content": "Render this chart spec as SVG:\n\n" + row["chart_spec"]},
            {"role": "assistant", "content": row["svg_code"]},
        ]
    }

Known limitations

• SVGs from synth-matplotlib carry matplotlib's stylistic defaults (font, axes, ticks). The model will produce matplotlib-flavored SVGs.
• svgen500k-* rows have no structured chart_spec -- only freeform title and _freeform_description are populated.
• SVGs are capped at 50 KB to keep training tractable.
• Only 6 chart types covered (bar, line, scatter, donut, histogram, area).

Citation

@dataset{regalado2026svgmix,
  author = {Regalado Cardoso, Daniel},
  title  = {SVG Chart Render Mix v1},
  year   = {2026},
  url    = {https://huggingface.co/datasets/DanielRegaladoCardoso/svg-chart-render-v1}
}

License

Pipeline and curation: Apache-2.0. nvBench data is MIT-licensed. svgen-500k rows carry per-row license fields from their upstream sources. See the source attribution table.

---

Built by Daniel Regalado Cardoso -- MSBA, University of Miami -- April 2026.