Real data sources, model hubs, and reference archives used throughout the AI learning path. Use these when experimenting, building projects, or going deeper on any topic.
How to use this page: Each section matches a stage in your roadmap. Classic ML datasets (UCI etc.) are great for understanding concepts hands-on. LLM pre-training datasets are useful for understanding scale. Model hubs let you download and run real models locally.
No resources have all of the selected labels. Clear a filter to see more.
🗄️ Classic ML & Tabular Datasets — Area 00–01
🏛️
UCI Machine Learning Repository
Classic datasets · Since 1987
⭐ Recommended Starting Point
The original open dataset archive — 650+ real-world datasets covering classification, regression, clustering and more. Iris, Wine, Boston Housing, Heart Disease, Adult Income are all here. Perfect for understanding vectors, embeddings and model training before scaling to LLMs.
ClassificationRegressionClustering650+ datasetsFree · No login
Massive community-driven dataset hub. Great for real competition data, notebooks, and seeing how others approach the same dataset. Includes tabular, image, NLP, and time-series.
All 700 verses of the Bhagavad Gita with chapter, verse number, Sanskrit original, transliteration, and English translation. A structured, clean, meaningful text corpus — excellent for practising tokenization, building embeddings, fine-tuning a small model, or building a RAG-based Q&A system over ancient wisdom.
💡 What you can build with it: semantic search over verses · verse recommendation · translation comparison · embeddings of concepts like dharma, karma, moksha
120k news headlines across 4 categories (World, Sports, Business, Sci/Tech). A go-to benchmark for text classification. Great for understanding how models learn to distinguish topics via embeddings.
Stanford Question Answering Dataset — 150k+ question/answer pairs from Wikipedia paragraphs. The standard benchmark for reading comprehension. Excellent for understanding how RAG retrieval + generation works.
The raw web crawl data behind GPT-3, LLaMA, and most large models. Petabytes of text scraped from the open web since 2008. Too large to download in full — but understanding it explains why LLMs know so much.
825GB curated open-source training corpus combining 22 high-quality datasets: Books3, GitHub, ArXiv, Wikipedia, StackExchange and more. Used to train GPT-Neo, GPT-J.
HuggingFace's own 15-trillion-token cleaned web dataset. Outperforms most existing open pre-training datasets on benchmarks. The current state-of-the-art open training corpus.
The world's largest open model hub. Download any model with one line. Includes BERT, GPT-2, LLaMA, Mistral, Falcon, Phi and thousands more in every size from 100M to 70B.
Quantized versions of large models that run on laptops without a GPU. A 7B model at 4-bit (Q4_K_M) needs only ~4GB RAM. Browse on Hugging Face and run via Ollama or llama.cpp.
Every major AI paper linked with its code, datasets, and benchmark results. See which model is state-of-the-art on any task. The best way to stay current with research.
Where all new AI research is published first. Filter by cs.LG (machine learning) or cs.CL (computational linguistics / NLP). Every major model — GPT, BERT, LLaMA — was first published here.
Free Jupyter notebooks with GPU access. Run HuggingFace models, fine-tune small models, experiment with tokenizers — all in-browser without installing anything locally.
Curated YouTube playlist covering foundational and frontier AI topics with the industry's best geometric, architectural, and mathematical visualizations. Master transformers, token embeddings, self-attention, and training dynamics visually.
The complete "If you…" chain — watch a sentence become tokens, embeddings, queries, keys, values, attention weights, context vectors, Transformer blocks, and finally an AI assistant. Type any sentence and click any token to see its live attention connections.
Tokens → EmbeddingsQKV AttentionSoftmax weightsTransformer blocksText generationRLHFLocal · no server
What a dimension is, how a token ID selects a row of the embedding matrix, why vocabulary size is not the embedding dimension, and how vectors are stacked into the Transformer input matrix.
The same 19-concept chain as an animated zigzag timeline — each step slides in from alternating sides with live visualizations for tokens, attention scores, probability bars, and generated text.
Animated Claude Code-style agent loop — watch the Think→Act→Observe cycle live with a streaming log, context window bar, tool highlights, and animated SVG orbit ring.