AI Systems · advanced
Training & Debugging
Init, NaNs, overfit tests, gradient checks.
Mental model
A training loop is forward pass → loss → backward pass → optimizer step, repeated over batches of data. When training looks broken, the data is the usual culprit (leaked, mislabeled, wrong shape) — check it before blaming the model.
How to study Training & Debugging
Begin by restating the mental model in your own words, then connect it to a concrete system you have built or operated. Name the mechanism, the constraint it addresses, and the trade-off it introduces. Use Stanford CS336 — L2 PyTorch & resource accounting (Spring 2025), CS336 Assignment 2 — Systems (profile, Triton, distributed training), A Recipe for Training Neural Networks (Karpathy) to check details, but close the source before writing your explanation. Retrieval is the learning step; rereading is only preparation.
Next, compare Training & Debugging with Checkpointing, LoRA & PEFT. Ask what changes in correctness, latency, resource use, operability, and failure recovery. Complete Overfit a single batch sanity check and preserve the command, input, output, and one failed attempt as evidence. Finish by explaining the idea without jargon to someone who has not studied the track.
Proof of understanding
- Explain the mechanism from first principles and identify the state it reads or changes.
- Give one situation where the concept is the right choice and one where it is not.
- Predict a realistic failure mode before running the drill, then compare the prediction with evidence.
- Connect the result to a roadmap or build artifact instead of treating the concept as isolated trivia.
Learn from primary sources
- Stanford CS336 — L2 PyTorch & resource accounting (Spring 2025) (course)
- CS336 Assignment 2 — Systems (profile, Triton, distributed training) (doc)
- A Recipe for Training Neural Networks (Karpathy) (article)
- The Illustrated GPT-2 — training a language model (article)
- Let's reproduce GPT-2 (124M) — Karpathy (video)
- Stanford CS336 — L9–11 Scaling laws (Spring 2025) (doc)
- How to Scale Your Model (JAX ML scaling book) (course)
Practice and explain it back
Overfit a single batch sanity check
Training loss stuck high on 1 batch of 8 examples. Name two data bugs to check before changing the architecture.
Expected evidence: Wrong labels/shapes, train-eval leakage, broken tokenizer, masked tokens all padding, etc.
Open the interactive drill →Review prompts
- What is the single fastest sanity check that your training loop is wired correctly?
Build evidence
Synthesize: AI Models & Training
Move from transformer foundations through pre-training, fine-tuning, post-training, compression, and evaluation. Produce one working system, benchmark, or evidence-backed design that integrates the path.
- Implements or precisely models the core mechanisms from all three milestones
- Includes at least one injected failure or adversarial case and demonstrates recovery
- Reports quality, latency, resource, reliability, or usability measurements relevant to the domain
- Ships a concise architecture note explaining decisions, trade-offs, and remaining risks