Optimizer steps (what matters)23,842 / 23,842
23,842
Progress bar shows (micro-steps)47,684 / 47,684
47,684
8
per_device_batch_size
2
gradient_accumulation_steps
23,842
max_steps (optimizer)
47,684
micro-steps (progress bar)
23,842 opt steps × 2 accum = 47,684 micro-steps — same training run, different counters