Hugging Face's logo

Datasets:
datapointai
/
text-to-svg-human-preferences-352k

Tasks:
Text-to-Image
Text Generation
Reinforcement Learning
Modalities:
Image
Tabular
Text
Formats:
parquet
Languages:
English
Size:
100K - 1M
Tags:
svg
vector-graphics
text-to-svg
code-generation
human-feedback
human-preferences
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

You agree to use this dataset under the CC-BY-4.0 license with attribution to Datapoint AI, and you agree not to attempt to re-identify annotators. Usage rights for the SVGs are governed by each model provider's terms.

Log in or Sign Up to review the conditions and access this dataset content.

Text-to-SVG human preferences: 352K votes across 14 models

This dataset contains the complete voting record behind the Datapoint SVG Bench leaderboard: 352,270 validated pairwise votes, exactly 10 for each of 35,227 SVG pairs. The votes compare 14 language models writing SVG code for 400 original prompts in 8 categories, judged by 20,323 annotators in 183 countries. Every output is included as SVG source, with the PNG annotators saw.

Built on the Datapoint annotation platform.

Key features

  • Ten judges on every pair. Each pair was judged by 10 people, so every comparison carries a preference margin rather than a single label.
  • Every model against every other model on every prompt. All 91 model pairs were scheduled on each prompt. A pair where one side produced no usable SVG was not sent to annotators (1,150 of 36,377).
  • Rendered the way a browser shows it. Chromium drew each SVG as an image at 1024×1024 px on white. Annotators saw two renders and the prompt, with one question: "Which image do you prefer?" Model names were never shown.
  • Source, render and cost for every output. The svgs config holds the SVG markup, the PNG, whether it rendered, and the tokens, cost and latency of the API call, including the 92 outputs that did not render.
  • Original prompts with checkable claims. 400 prompts written for this benchmark, 50 per category, with 796 claims such as "exactly four icons in order: previous, play, pause, next".

Categories

Category What it stresses Prompts
icons geometry, consistent stroke, set consistency 50
logos exact brand text, mark and lettering together, negative space 50
illustrations characters, animals, objects, props, counts 50
scenes composition, depth, spatial relations between named objects 50
diagrams stated data, labels, flow logic (charts, flowcharts, maps) 50
typography exact strings, hierarchy, text on paths, lettering effects 50
ui screens and components with exact labels, values and states 50
precision counts, clock times, geometry, spatial reasoning 50

Preview

The same held-out prompt, drawn by four of the models:

A lighthouse on a rocky cliff at sunset, with three seagulls flying and waves crashing below.

GPT-6 Astra
GPT-6 Astra		 Arrow 2 Telos
Arrow 2 Telos		 Gemini 3.8 Flash
Gemini 3.8 Flash		 Claude Opus 5.5
Claude Opus 5.5

Dataset structure

Config Rows Description
svgs (default) 5,600 One output per model and prompt: SVG source, the rendered PNG, and generation metadata
pairs 35,227 One SVG pair with vote counts, labels, and the winner
responses 352,270 One individual human vote
prompts 400 One prompt with its category and checkable claims
models 14 One model with its API identifier, settings, validity, and cost

svgs

Each row is one model output. The PNG annotators saw is embedded in the image column, so the Dataset Viewer shows it directly.

Column Type Description
image image The 1024×1024 PNG annotators saw. Null when the SVG did not render
svg string SVG markup taken from the model's answer. Null when the answer had none
svg_key string {model_id}/{prompt_id} join key used by pairs
model_id, prompt_id, category string Model, prompt, and category
valid bool The SVG rendered and was shown to annotators
status string ok (5,508), xml_error (42, not well-formed XML), no_svg (41, no SVG in the answer), or truncated (9, cut off at the output limit)
finish_reason string Finish reason reported by the API
input_tokens, output_tokens, reasoning_tokens int Token usage reported by the API
cost_usd float Cost of the call at list prices, from the reported usage
latency_seconds float Wall time of the API call
svg_bytes, svg_elements int Size of the SVG and its number of elements
image_width, image_height, image_sha256 mixed Render size and the SHA-256 digest of the PNG

pairs

Each row is one comparison. svg_a_key and svg_b_key join to svgs.

Column Type Description
pair_key string Unique pair identifier
prompt_id, category string Prompt and its category
model_a, model_b string The two models being compared
svg_a_key, svg_b_key string Join keys into the svgs config
votes_a, votes_b int Vote counts used by the leaderboard; they sum to 10
label_a, label_b float Preference fractions. Ties are 0.5/0.5. Use these directly for DPO or reward-model training
trust_weighted_votes_a, trust_weighted_votes_b float Vote counts weighted by trust score
winner string a, b, or tie
num_votes int Total votes for the pair (10)

responses

Each row is one human vote. The config holds exactly the 10 validated votes per pair that the leaderboard counts.

Column Type Description
pair_key string Joins to pairs
prompt_id, category string Prompt and its category
chosen string a or b, the image the annotator preferred
annotator string Salted hash. Stable across the dataset, not linkable to accounts
trust_score float Annotator trust score (0 to 1) when the vote was cast
time_taken_ms int Time spent on the judgment. The median is about 9 seconds
completed_at timestamp When the vote was cast
country string Annotator country (ISO 3166-1 alpha-2)

prompts

Column Type Description
prompt_id, category string Prompt identity and category
prompt string The prompt sent to the models and shown to annotators
claims list Facts a correct SVG should show. check is auto when the claim can be checked from the image or the SVG text, and human when it needs a person
num_claims int Number of claims (1 or 2)

models

Column Type Description
model_id, display_name, developer string Model identity
api_model, reasoning_setting, max_output_tokens mixed The API model and the generation settings used
num_prompts, valid_svgs int Prompts attempted and SVGs that rendered
total_cost_usd, median_latency_seconds float Cost of all 400 calls at list prices, and median call time

Splits

The svgs, pairs, responses, and prompts configs have train and test splits. The test split holds out 40 prompts, 5 per category, chosen with a fixed seed. No prompt or SVG appears in both splits, so you can train a reward model on train and evaluate it on test without prompt leakage. models is a reference table.

Usage 

from datasets import load_dataset

pairs = load_dataset("datapointai/text-to-svg-human-preferences-352k", "pairs", split="train")
row = pairs[0]
print(row["prompt_id"], row["model_a"], "vs", row["model_b"], "->", row["winner"], f'({row["votes_a"]}-{row["votes_b"]})')

Use for DPO training

With 10 votes per pair, you can select training pairs by preference margin. Filtering to decisive pairs (for example, 7–3 or stronger) removes near-tie label noise:

svgs = load_dataset("datapointai/text-to-svg-human-preferences-352k", "svgs", split="train").remove_columns("image")
code = dict(zip(svgs["svg_key"], svgs["svg"]))
prompts = {p["prompt_id"]: p["prompt"] for p in load_dataset("datapointai/text-to-svg-human-preferences-352k", "prompts", split="train")}
MIN_MARGIN = 0.7  # keep pairs where the winner took >= 7 of 10 votes

def to_dpo(row):
    confidence = max(row["label_a"], row["label_b"])
    if row["winner"] == "tie" or confidence < MIN_MARGIN:
        return None
    chosen, rejected = ("a", "b") if row["winner"] == "a" else ("b", "a")
    return {
        "prompt": prompts[row["prompt_id"]],
        "chosen": code[row[f"svg_{chosen}_key"]],
        "rejected": code[row[f"svg_{rejected}_key"]],
        "confidence": confidence,
    }

dpo_rows = [d for d in (to_dpo(r) for r in pairs) if d]

Look at the renders 

svgs = load_dataset("datapointai/text-to-svg-human-preferences-352k", "svgs", split="test")
row = next(r for r in svgs if r["valid"])
row["image"].save("render.png")  # PIL image, 1024x1024
print(row["svg"][:200])

Work with individual votes 

# Every vote carries its timing and the annotator's country.
responses = load_dataset("datapointai/text-to-svg-human-preferences-352k", "responses", split="train")
considered = responses.filter(lambda r: r["time_taken_ms"] >= 5_000)  # votes that took at least 5 seconds

Leaderboard

Elo ratings from a Bradley–Terry fit on raw votes, centered so the 14 models average 1000. The Overall board gives each of the eight categories an equal share of the fit and matches the live leaderboard. Rank spread shows the ranks each model can hold within its 95% confidence interval. Only voted pairs enter the fit.

SVG model Elo rankings: 14 models ranked by Elo score
Rank Model Developer Elo 95% CI Rank spread Win rate
1 GPT-6 Astra OpenAI 1065.6 1059–1072 1–2 59.8%
2 Gemini 3.8 Flash Google 1059.2 1052–1066 1–3 58.8%
3 Muse Wonder Meta 1047.3 1042–1053 2–3 57.1%
4 Claude Fable 5.1 Anthropic 1028.5 1023–1034 4–7 54.3%
5 Claude Opus 5.5 Anthropic 1028.2 1023–1034 4–7 54.3%
6 Qwen3.8 Max Alibaba 1018.4 1013–1024 4–8 52.7%
7 Gemini 3.1 Pro Google 1016.9 1011–1023 4–8 52.5%
8 Arrow 2 Telos QuiverAI 1011.6 1004–1020 6–9 51.7%
9 GLM-5.3 Z.ai 1002.6 997–1008 8–9 50.3%
10 GPT-6 Luna OpenAI 971.8 966–978 10–11 45.7%
11 Grok 4.7 xAI 968.7 963–975 10–11 45.2%
12 Kimi K3 Moonshot AI 956.6 950–963 12–12 43.4%
13 MiniMax M3 MiniMax 926.7 920–934 13–13 39.0%
14 DeepSeek V4 Pro DeepSeek 897.8 891–904 14–14 34.8%

Category leaders:

Category 1st 2nd 3rd
Icons Claude Opus 5.5 (1070) Gemini 3.8 Flash (1063) GPT-6 Astra (1060)
Logos Gemini 3.8 Flash (1078) Muse Wonder (1052) GPT-6 Astra (1047)
Illustrations Arrow 2 Telos (1128) GPT-6 Astra (1113) Claude Fable 5.1 (1050)
Scenes Arrow 2 Telos (1133) GPT-6 Astra (1121) Qwen3.8 Max (1052)
Diagrams Gemini 3.8 Flash (1092) Muse Wonder (1070) GPT-6 Astra (1048)
Typography GPT-6 Astra (1057) Gemini 3.8 Flash (1056) Muse Wonder (1045)
UI Gemini 3.8 Flash (1061) Muse Wonder (1050) GPT-6 Astra (1047)
Precision Gemini 3.8 Flash (1057) Muse Wonder (1046) GPT-6 Astra (1042)

Rendering and cost, from the svgs and models configs. Cost is the list-price cost of all 400 calls divided by the SVGs that rendered:

Model SVGs that rendered Cost per rendered SVG Median call time
GPT-6 Astra 400 / 400 $0.260 87 s
Gemini 3.8 Flash 369 / 400 $0.112 180 s
Muse Wonder 398 / 400 $0.159 79 s
Claude Opus 5.5 400 / 400 $0.051 21 s
Claude Fable 5.1 400 / 400 $0.276 48 s
Qwen3.8 Max 397 / 400 $0.132 447 s
Gemini 3.1 Pro 400 / 400 $0.254 147 s
Arrow 2 Telos 400 / 400 $0.329 117 s
GLM-5.3 388 / 400 $0.079 114 s
GPT-6 Luna 399 / 400 $0.003 49 s
Grok 4.7 400 / 400 $0.175 377 s
Kimi K3 399 / 400 $0.051 33 s
MiniMax M3 364 / 400 $0.032 116 s
DeepSeek V4 Pro 394 / 400 $0.043 147 s

Comparison to related work

This dataset Rapidata SVG Benchmark
Human judgments 352,270 votes 3,406,090 responses
Comparisons 35,227 517,590
Questions 1 (preference, prompt shown) 3 (preference, coherence, alignment)
Votes per comparison 10 varies
Models 14 67
Prompts 400, written for the benchmark, 8 categories 500, about 450 drawn from chat logs
Every model pair on every prompt Yes No
Renderer Chromium, 1024 px cairosvg, 768 px
Cost and latency per SVG Yes No

Rapidata's figures are as stated on its card on 2026-10-15.

Intended use

Use this dataset to:

  • Train and evaluate reward or preference models for SVG and code-based image generation, including DPO, RLHF, and best-of-N reranking.
  • Compare models on quality, rendering reliability, and cost together.
  • Study inter-annotator agreement and annotation quality at scale.
  • Audit the published leaderboard.

The SVGs are one output per model and prompt, collected for evaluation. They are not a curated training corpus for SVG generation. Don't use this dataset to attempt to identify annotators.

License

Votes, prompts, and all metadata are released under CC-BY-4.0. The SVGs and their renders are outputs of the listed third-party models and are distributed for research and evaluation. Usage rights for model outputs are governed by each provider's terms.

More Datapoint datasets

Citation 

@dataset{datapoint_svg_preferences_2026,
  title   = {Text-to-SVG Human Preferences 352K: the Datapoint SVG Bench voting record},
  author  = {{Datapoint AI}},
  year    = {2026},
  url     = {https://huggingface.co/datasets/datapointai/text-to-svg-human-preferences-352k},
  note    = {352,270 pairwise votes over 14 models and 400 prompts, 10 votes per pair}
}
Downloads last month
173
Total file size:
1.22 GB