Datasets & Tool Archive

In one sentence: a practical archive answering "where to get data and what tools to use" — classic datasets grouped by modality (size, task, source, notes) + one-line profiles of common tools + decision principles for choosing datasets.

Timeliness note: Dataset sizes, versions, and other information are current as of 2026-08. Verify before citing. Download links all point to official channels.

Data is the raw material of deep learning: the model's ceiling is determined by the data. See Data & Data Engineering. This page addresses three practical questions: which dataset to use for practice, where to download, and what tools to run with.

1. Image Datasets

Dataset	Size	Task	How to Get	Notes
MNIST	70K 28×28 handwritten digits	10-class classification	Official site / direct download via Torchvision	Too "clean" — you can't overfit on it. Good for running through a pipeline, not for chasing SOTA scores
CIFAR-10/100	60K 32×32 color images	10/100-class classification	Official site / Torchvision	Low resolution, but the standard battleground for quickly validating network architectures
ImageNet	~14M images, 22K classes	Image classification, pre-training benchmark	Official site (apply for download)	Large volume (hundreds of GB). Academic use requires agreeing to terms. Most people do pre-training comparisons on ImageNet-1K
COCO	330K images, 2.5M annotated instances	Detection / segmentation / keypoints / captions	Download files from official site	Annotation format (JSON) needs parsing. TorchVision has built-in loaders
Open Images	9M+ images	Detection / segmentation / visual relationships	Google Cloud storage shards	Large scale, higher annotation noise than COCO, suited for pre-training

Tip: For all image datasets, first run through the pipeline using TorchVision's built-in interfaces before considering full downloads. When processing large-scale data, use DataLoader multi-processing and caching. See Training Recipes & Hyperparameter Tuning for details.

2. Text Datasets

Dataset	Size	Task	How to Get	Notes
GLUE	A suite of 9 NLU tasks (includes SST-2, MNLI, QQP, etc.)	Sentence classification, natural language inference, similarity, etc.	HuggingFace datasets	For Chinese, check out CLUE as a counterpart. A classic benchmark for general language capability
SuperGLUE	A harder version of GLUE (8 more difficult tasks)	Reasoning, common sense, etc.	HuggingFace datasets	Fewer samples — pay attention to metric computation details
SQuAD	100K+ QA pairs (SQuAD1.1); includes unanswerable questions (SQuAD2.0)	Extractive QA	Official site / HuggingFace	Version 2.0 adds unanswerable samples, testing the model's ability to "refuse to answer"
Wikipedia (English)	Billions of tokens	Pre-training corpus, RAG knowledge source	Official dumps, released monthly	Enormous volume — usually download subsets on demand. Commonly used as a knowledge base for RAG practice. See Large Language Models (LLMs)
Common Crawl	Hundreds of TB of web data	Large-scale pre-training corpus	Official monthly snapshots	Requires extensive cleaning and deduplication. Hard for individuals to use in full — usually use subsets
The Pile	820GB multi-source text	Open-source pre-training corpus	Official site / HuggingFace	Diverse components (papers, GitHub, books, conversations, etc.). A common corpus for large-scale pre-training research

Tip: Chinese corpora can be obtained from HuggingFace's Chinese datasets (e.g., C4 Chinese version, CLUECorpus). Any pre-training corpus needs deduplication, filtering, and privacy cleaning. See Data & Data Engineering.

3. Speech & Audio Datasets

Dataset	Size	Task	How to Get	Notes
LibriSpeech	~1,000 hours of English read speech	Automatic Speech Recognition (ASR)	OpenSLR shard downloads	16kHz sampling rate, high-quality annotation alignment — the standard ASR benchmark
Common Voice	Multilingual crowdsourced speech (continuously growing)	ASR, multilingual	Mozilla official download	High noise, diverse accents — realistic but uneven annotation quality. See Speech & Audio

4. Multimodal & Molecular/Graph Datasets

Dataset	Size	Task	How to Get	Notes
LAION-5B	5.8B image-text pairs	Image-text pre-training, CLIP training	Official shard-index downloads	Unfiltered crawled data — contains harmful/low-quality content. Research use requires attention to compliance and content review
Conceptual Captions (CC)	~3.3M image-text pairs	Image-text pre-training	GitHub official link	Requires downloading original images separately; link validity needs self-checking
QM9	134K organic molecules	Molecular property regression	Official site	A standard entry-level dataset for graph/molecular ML. See Graph Neural Networks
OGB (Open Graph Benchmark)	Multi-task graph benchmarks (node / edge / graph-level)	Graph learning benchmark	Official site / HuggingFace	Has official evaluation protocols and leaderboards. Don't tune on the test set

5. Tool Profiles

Tool	One-line Description	Use Case
PyTorch	The dominant deep learning framework	Main modeling and training framework
TorchVision	PyTorch's official vision toolkit	Image dataset loading, pre-trained models, transforms
Hugging Face Transformers	Unified API for loading thousands of pre-trained models	Text/vision/speech model loading and fine-tuning
Hugging Face datasets	Dataset loading and processing library	Stream large corpora, unified preprocessing
diffusers	Official diffusion model toolkit	Generative model training and inference. See Diffusion Models & Generative AI
peft	Parameter-efficient fine-tuning toolkit	LoRA and other fine-tuning methods out of the box
accelerate	Distributed training simplifier	Run multi-GPU / mixed-precision training with a few lines of code
ONNX Runtime	Cross-platform inference engine	Model export and deployment. See MLOps & Model Deployment
TensorBoard / W&B	Visualization and experiment tracking	Metric monitoring, experiment comparison
OpenAI CLIP	Image-text contrastive learning model	Image-text retrieval, zero-shot classification, feature extraction

Tool selection advice

Beginners should use the "PyTorch + HuggingFace family (Transformers/datasets/peft/accelerate)" end-to-end. For research efficiency, learn JAX. Add ONNX / inference frameworks for production deployment. A more complete selection methodology is at How to Choose Frameworks & Tools.

6. Principles for Choosing Datasets

Principle 1: Task Before Data

First clarify "what problem to solve and what metric to use for evaluation," then choose data. If the metric is wrong, no dataset size matters. See Deep Learning Evaluation & Experimentation.

Principle 2: Size Matches the Stage

• Learning principles / running through a pipeline: MNIST, CIFAR — validate ideas within minutes;
• Serious projects / portfolio: COCO, GLUE, LibriSpeech — scale and annotation quality sufficient for resume writing;
• Pre-training / research: ImageNet, The Pile, LAION — requires compute and engineering support.

Principle 3: Watch for Three Categories of Data Problems

1. Leakage: Training/test mixing, temporal leakage of future information — inflates metrics. See Common Pitfalls & Anti-patterns;
2. Distribution: Dataset distribution differs from the real-world scenario — the model fails upon deployment. See Evaluation in Practice;
3. Compliance: Licensing and ethical issues for crawled data, person images, and copyrighted text — especially critical for datasets like LAION. See Interpretability & Fairness.

Principle 4: Prioritize "Ecosystem-Rich" Datasets

Datasets with official loaders (TorchVision / HF datasets), public baseline scores, and community discussion make your results comparable and interpretable — which is why ImageNet and GLUE have become de facto standards.

Further Reading

• Data & Data Engineering — Data cleaning, annotation, and imbalance handling
• Training Recipes & Hyperparameter Tuning — Data loading and training pipeline optimization
• Deep Learning Evaluation & Experimentation — Choose the right metrics before choosing data
• Large Language Models (LLMs) — Pre-training corpora and RAG data practices
• How to Choose Frameworks & Tools — Toolchain selection methodology
• Portfolio Projects — Use datasets to build projects worth writing on your resume

References

• Papers with Code. Datasets (dataset search and benchmarks)
• Hugging Face Datasets Official Documentation (continuously updated)
• TorchVision Datasets Official Documentation (continuously updated)