datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
javanese-Komodo-pixelgpt
Javanese PixelGPT Dataset
This dataset contains preprocessed Javanese text data for training PixelGPT models.
Dataset Statistics
Language: Javanese (jawa)
Total samples: 401,542
Train samples: 400,726
Test samples: 816
Tokenizers
[TO BE EDITED]
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel representation of aksara text
<tokenizer_name>_token_ids: Token IDs from grapheme-based… See the full description on the dataset page: https://huggingface.co/datasets/Exqrch/javanese-Komodo-pixelgpt.pixel-art-bench-v1
Pixel Art Benchmark Dataset (Source)
The Pixel Art Benchmark Dataset is a structured collection of pixel-art outputs generated by large language models (LLMs). Each sample consists of a discrete color palette and a grid-based representation of pixel art, along with generation metadata such as token usage, cost, and model provenance.
Each row in the dataset represents a single generated pixel-art sample.
Encoding Details
Each string in grid represents one row of pixels.… See the full description on the dataset page: https://huggingface.co/datasets/AINovice2005/pixel-art-bench-v1.llm-smartrouter-benchmark
LLM SmartRouter & Agent Highway Latency & Cost Benchmark (v1.4.0)
Empirical performance benchmark dataset comparing direct model endpoints (OpenAI, Anthropic Claude, Google Gemini) against the PixelRouter / BLUN SmartRouter proxy layer and Autonomous Agent Web Highway (https://api.pixeloffice.eu/v1).
v1.4.0 Benchmark Highlights
Anthropic Claude Messages API: Sub-35ms proxy routing for native /v1/messages payloads with 94%+ cost savings.
Machine Web Highway… See the full description on the dataset page: https://huggingface.co/datasets/pixeloffice/llm-smartrouter-benchmark.balinese-Komodo-pixelgpt
Balinese PixelGPT Dataset
This dataset contains preprocessed Balinese text data for training PixelGPT models.
Dataset Statistics
Language: Balinese (bali)
Total samples: 54,467
Train samples: 54,017
Test samples: 450
Tokenizers
[TO BE EDITED]
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel representation of aksara text
<tokenizer_name>_token_ids: Token IDs from grapheme-based… See the full description on the dataset page: https://huggingface.co/datasets/Exqrch/balinese-Komodo-pixelgpt.sundanese-pixelgpt
Sundanese PixelGPT Dataset
This dataset contains preprocessed Sundanese text data for training PixelGPT models.
Dataset Statistics
Language: Sundanese (sunda)
Total samples: 294,756
Train samples: 293,933
Test samples: 823
Tokenizers
Grapheme tokenizer: izzako/sunda-llama-tokenizer
LLaMA tokenizer: ernie-research/DualGPT
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel representation… See the full description on the dataset page: https://huggingface.co/datasets/izzako/sundanese-pixelgpt.balinese-pixelgpt
Balinese PixelGPT Dataset
This dataset contains preprocessed Balinese text data for training PixelGPT models.
Dataset Statistics
Language: Balinese (bali)
Total samples: 54,467
Train samples: 54,017
Test samples: 450
Tokenizers
Grapheme tokenizer: izzako/javanese-llama-tokenizer
LLaMA tokenizer: ernie-research/DualGPT
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel… See the full description on the dataset page: https://huggingface.co/datasets/izzako/balinese-pixelgpt.pinchbench-sclean-trajectories
PinchBench Lean Trajectories
Full agent trajectories and scores for small models run through a stripped down,
single model opencode backend on the 116 task PinchBench suite. One row per
(model, task, run).
Harness
The backend is a lean, three agent opencode setup, not the stock product stack:
default agent: the coordinator. Talks to the task, does simple work with
its own tools (read, edit, bash, todowrite), and delegates the rest.
websearch-agent: the only path… See the full description on the dataset page: https://huggingface.co/datasets/pixelxiong/pinchbench-sclean-trajectories.aws-malay-qa
AWS Q&A in Bahasa Melayu (aws-malay-qa)
~2.6k AWS question–answer pairs in Bahasa Melayu, in chat messages format, used to train
PixelSpaceAI/Malaysian-Qwen2.5-7B-AWS-Malay-LoRA.
Technical terms are kept in English (S3, Lambda, IAM, bucket, policy) the way Malaysian engineers speak.
Files
File
Rows
Purpose
train.jsonl
2,511
training split
eval.jsonl
150
held-out eval, service-stratified (never in train)
Schema
{
"source":… See the full description on the dataset page: https://huggingface.co/datasets/PixelSpaceAI/aws-malay-qa.pixel-art-bench-lite
🎨 Pixel Art Bench Lite
Pixel Art Bench Lite is a structured-output benchmark designed to evaluate small language models on their ability to generate valid, interpretable, and semantically meaningful JSON outputs under strict constraints.
The benchmark is based on Pixel Art Bench focuses on pixel art generation over a fixed 24×24 grid, requiring models to produce outputs that are syntactically correct but also visually coherent.
While many benchmarks evaluate free-form text… See the full description on the dataset page: https://huggingface.co/datasets/AINovice2005/pixel-art-bench-lite.sft-aeo-telemetry-dataset
SFT AEO & AI Crawler Telemetry Instruction Dataset (2,100 Samples)
Curated, high-precision Supervised Fine-Tuning (SFT) dataset containing 2,100 instruction-following pairs formatted in standard ChatML / OpenAI JSONL.
Published by Pixel Office EU.
Core Dataset Domains (2,100 Samples):
Showcase Architecture & MCP Tool Specifications (780 samples): 195 verified B2B software architectures with Model Context Protocol (MCP) schemas and sub-35ms edge latency… See the full description on the dataset page: https://huggingface.co/datasets/pixeloffice/sft-aeo-telemetry-dataset.llm-rag-optimized-schema-templates
Schema.org JSON-LD Templates Optimized for LLM RAG Retrieval (2026)
Curated dataset of Schema.org JSON-LD templates designed, tested, and optimized for Retrieval-Augmented Generation (RAG) systems, SearchGPT, Gemini, and Claude search parsers.
Published by Pixel Office EU.
Purpose
Standard Schema.org markup is often too nested or dense for token-efficient LLM context window ingestion. These templates prioritize high-salience fields that crawlers prioritize when… See the full description on the dataset page: https://huggingface.co/datasets/pixeloffice/llm-rag-optimized-schema-templates.pixels_vs_code
Pattern Over Pixels Screenshot-to-Code
This dataset contains controlled counterfactual screenshot-to-code examples built
from 30 real-world webpages from Design2Code. Each example preserves a repeated
UI pattern while introducing a single localized deviation, allowing researchers
to test whether multimodal models follow the pixels or simply restore the
dominant template.
Contents
720 perturbed HTML instances
360 structural-card examples
360 text-style examples
2… See the full description on the dataset page: https://huggingface.co/datasets/AnonymousSubmissionASE/pixels_vs_code.lampung-pixelgpt
Lampung PixelGPT Dataset
This dataset contains preprocessed Lampung text data for training PixelGPT models.
Dataset Statistics
Language: Lampung (lampung)
Total samples: 1,029
Train samples: 945
Test samples: 84
Tokenizers
Grapheme tokenizer: izzako/sunda-llama-tokenizer
LLaMA tokenizer: ernie-research/DualGPT
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel representation of aksara… See the full description on the dataset page: https://huggingface.co/datasets/izzako/lampung-pixelgpt.javanese-pixelgpt
Javanese PixelGPT Dataset
This dataset contains preprocessed Javanese text data for training PixelGPT models.
Dataset Statistics
Language: Javanese (jawa)
Total samples: 401,542
Train samples: 400,726
Test samples: 816
Tokenizers
Grapheme tokenizer: izzako/javanese-llama-tokenizer
LLaMA tokenizer: ernie-research/DualGPT
Features
text_id: Document identifier
chunk_id: Chunk identifier within document
pixel_values: Rendered pixel… See the full description on the dataset page: https://huggingface.co/datasets/izzako/javanese-pixelgpt.pixelprose-sample-5k
Dataset Card for pixelprose-sample-5k-gemini
Dataset Description
This dataset contains 5,000 image captions from the PixelProse dataset, enhanced with AI-generated SVG representations and reasoning traces using Google's Gemini 2.5 Flash Preview model with thinking mode enabled(google/gemini-2.5-flash-preview:thinking).
Dataset Summary
Source: Sampled from tomg-group-umd/pixelprose
Size: 5,000 samples
Model Used:… See the full description on the dataset page: https://huggingface.co/datasets/thesantatitan/pixelprose-sample-5k.synth-dent-test