LabRift AI Academy

From Zero to AI Expert — rebuilt as real experiments
Academy v2 · Built by Vega · Labs 1–2 interactive
🍎 Apple Silicon Native🧪 Hands-On💰 Zero Cloud Cost🦝 Raccoon Approved

Your journey

Send feedback
0/10
0 modules complete10 modules7 hands-on labs planned$0 cloud cost

Start at Module 1. Your progress is stored only in this browser—no account or tracking yet.

A base model is a broad textbook. Fine-tuning adds a narrow set of reliable notes in the margins. LoRA (Low-Rank Adaptation) makes that practical by training a small adapter instead of changing every model weight.

What you will learn

  • Why prompting, RAG, and fine-tuning solve different problems
  • How a JSONL training example pairs an instruction with a desired answer
  • What an adapter is and why it is much smaller than the base model
  • How training loss is a useful signal—but not proof of intelligence
LAB 01 · BROWSER SIMULATION

Train a tiny adapter

This is an honest educational simulation of the fine-tuning loop. It runs instantly in the browser so testers do not need a 4B model. The companion Mac run will use the real Qwen3/MLX pipeline later.

1. Build training examples

2. Run the adapter training

The simulated run updates the adapter signal across 60 iterations. Watch the loss fall, then compare the base and adapted answers.

Ready
loss: —iteration 0/60

3. Compare the behavior

Base model · before adapterRun the lab to see the baseline response.
Adapted model · after adapterRun the lab to see the learned response.

No adapter has been trained yet.

Checkpoint

Which change is the best fit for fine-tuning?

What does LoRA usually train?

Attention is a learned way to decide which parts of a sequence matter when processing a particular token. The core calculation is simple enough to inspect: make queries and keys, compare them, then turn the scores into probabilities.

Important: this lab uses a small deterministic educational attention engine, not a pretrained language model. That keeps it fast, private, and understandable while still showing the real Q/K → score → softmax pipeline.
LAB 02 · LIVE ATTENTION VISUALIZER

Watch tokens attend to one another

Checkpoint

In the heatmap, each row represents…

What happens after raw attention scores are calculated?

Compare how different tokenizers split the same sentence, measure context usage, and learn why spelling and arithmetic can be surprisingly hard for language models.

Planned lab: a local tokenizer comparison with token counts, BPE-style merges, and context-window experiments.

Explore FP32, FP16, INT8, and 4-bit tradeoffs, including GGUF and MLX formats.

Planned lab: compare model size, memory, speed, and output quality with a controlled quantization demo.

Build a mini retrieval system and compare prompting, retrieval, and adaptation on the same question.

Planned lab: chunk, index, retrieve, and inspect the evidence used to answer a question.

Inspect the three-stage training story and compare base versus instruct behavior.

Planned lab: diagnose a model training configuration and identify which stage produced each behavior.

See why parameter count is not the same thing as compute, capability, or cost.

Planned lab: inspect model configs and trace a simplified mixture-of-experts routing decision.

Generate controlled outputs and learn what actually changes when temperature, top-p, and top-k move.

Planned lab: run a sampling playground and measure tokens per second with cache settings.

Build a model preference matrix around concrete tasks instead of hype.

Planned lab: benchmark three models on medical billing, creative writing, and reasoning.

Learn the factions, vocabulary, funding cycle, and strategic ideas shaping the AI industry.

Planned lab: analyze real claims and separate technical signal from culture-war noise.