@david-cortes Hello , I’ve updated test_basic.R to cover all the scenarios you requested.
The new tests now verify:

• Serialization: Saving and loading an nrounds=0 model preserves the state correctly (and keeps niter at 0).
• Continuation: I confirmed that initializing with 0 rounds and then continuing for 10 rounds produces numerically identical predictions to training 10 rounds from scratch (using fixed seeds and a strict tolerance).
• Callbacks: I verified that early_stopping and evals work fine with empty models.

One small detail: I noticed that when continuing training from an empty model with early stopping, the niter attribute sometimes stays NULL (likely because the callback can't find a previous "best iteration"). I updated the test to explicitly check that the predictions have diverged/updated in this case, which proves the training is working correctly even if the attribute is missing.

Let me know if there are any other changes to be made!
Thanks!!!

@trivialfis I have updated the tests to cover all the requested scenarios (serialization, continuation, and callbacks).

Note on Case 3 (Callbacks):
I noticed that the existing R wrapper has a separate issue where predict() crashes on 0-round models with early stopping (due to missing best_iteration). To work around this while still proving that continuation works, the test now validates that the continued model produces learned predictions (non-uniform standard deviation) rather than checking the niter attribute directly.
The R-package checks are passing now (the Mac/Python failure seems unrelated to these changes).

Ready for review.
Thanks!!!!

Thanks @david-cortes for pointing these out to me.

I have updated the code to:

• Set nthread = 1 in all xgb.train calls to satisfy CRAN requirements.
• Use xgb.save.raw for serialization testing.
• Reuse the global train variable instead of reloading the dataset.

Ready for review!
If anything else left let me know.
Thanks!!