Make Default Checkpointing Behavior Extremely Conservative

[Helw150]

For 99% of our runs, we don't really care about intermediate checkpoints, except for resumption. We should make it so that the default behavior (for default train and elsewhere) only keeps the final checkpoint. @claude

[claude]

Claude finished @Helw150's task in 4m 22s —— View job • claude/issue-4456-20260406-2314 • Create PR ➔

---

Task: Make default checkpointing conservative (only keep final)

• Investigate checkpointing defaults in experiments/defaults.py and Levanter's CheckpointerConfig
• Confirm the trainer forces a final checkpoint via run_hooks(info, force=True) at end-of-training (lib/levanter/src/levanter/trainer.py:559,572), so an empty keep list still produces a final permanent checkpoint
• Update defaults so default_train / default_sft / default_dpo keep no permanent intermediate checkpoints by default
• Run ./infra/pre-commit.py --all-files --fix — OK
• Run tests/test_training.py — 4 passed
• Commit & push

Summary

Changed steps_per_export (SimpleTrainConfig) and steps_per_checkpoint (SimpleSFTConfig, SimpleDPOConfig) defaults from a positive int to None. Added _checkpoint_keep() helper in experiments/defaults.py that maps None → keep=[], so the default CheckpointerConfig retains:

• a single rolling temporary checkpoint (every save_interval=10min) for resumption, and
• the final permanent checkpoint at end of training.

Experiments that explicitly set steps_per_export / steps_per_checkpoint are unaffected.

Testing

• ./infra/pre-commit.py --all-files --fix: OK
• uv run --with pytest --with pytest-timeout python -m pytest tests/test_training.py -x -q: 4 passed

· branch claude/issue-4456-20260406-2314