Step versus epoch in a training run: define the relationship.

Attempt it

A step is exactly one optimizer update, the model has done forward + backward + optimizer.step() once, over one effective batch. An epoch is one full pass through the training dataset. The two are linked by the formula:

steps_per_epoch = ceil(num_examples / effective_batch_size)

where effective_batch_size = micro_batch x grad_accum_steps x num_gpus. So '3 epochs at effective batch 32 on 12,800 examples' resolves to 3 x ceil(12800 / 32) = 1,200 steps. Trainer configs let you cap a run by max_steps OR num_train_epochs; whichever ends first wins.

Your answer

Quick reference

TL;DR

A step is one optimizer update over one effective batch; an epoch is one full pass through the dataset. Steps per epoch equals dataset size divided by effective batch size.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Think of reading a thick novel. A step is reading one chunk of pages and pausing to think about what you just read. One full pass through the whole book is the bigger unit. If the book has 1,200 pages and you read 32 pages between pauses, you pause 1,200 divided by 32 times to finish the book once. That is the relationship. The page count is how much data you have, the pause cadence is the chunk size you set, and the number of pauses is how many small updates fit inside one full read-through. A trainer can stop when you finish three books or after a fixed number of pauses, whichever comes first.

Key concepts

Training loop Gradient accumulation Batch size

Quick scan · 2–4 min

Everything important, quickly.

3 min: define step and epoch, give the steps_per_epoch formula, name the three factors in the effective batch, and end on tokens seen as the portable budget unit.

Side-by-side

Term	Counts what	Depends on
Step	One optimizer update	Effective batch size
Epoch	One full data pass	Dataset size
Effective batch	Examples per update	Micro-batch, accumulation, GPU count
Tokens seen	Total training work	Examples processed times sequence length

Real-world examples

Real products, models, and research that use this idea.

Hugging Face Trainer exposes both num_train_epochs and max_steps, and 2026 Llama 4 Maverick fine-tune recipes typically pin max_steps for predictable budgets.
Axolotl YAML configs print the resolved steps per epoch on startup so teams can sanity check before burning GPU time.
DeepSpeed and FSDP both compute effective batch from the world size, so a Claude Opus 4.7 distillation run that moves from 8 to 32 GPUs sees quartered steps per epoch at the same micro-batch.
WandB and Comet dashboards log step on the x-axis by default, which is why teams now add tokens-seen as a secondary axis to compare runs across cluster shapes.
Unsloth tutorials emphasise that doubling grad_accum_steps halves the step count for the same training data, a frequent source of confusion in beginner threads.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QIf you double grad_accum_steps with the same micro-batch and GPU count, what happens to steps per epoch and to the learning rate you should use?

Sign in or go Pro to see the approach for this follow-up.

QWhy do production teams log tokens seen as the primary budget axis instead of steps?

Sign in or go Pro to see the approach for this follow-up.

QHow do warmup schedules interact with the step versus epoch choice?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.