How to debug histogram bins, candidate split points, and gain values during training?

[thoongee]

Hello, I am a student studying machine learning in Korea. I am currently using the XGBoost library in a Python environment.

I used the following code to inspect the trained model:

booster = xgb_model.get_booster()
dump_list = booster.get_dump(dump_format='json')
print("\n=== Official XGBoost Tree Dump ===")
for i, tree in enumerate(dump_list):
    print(f"Tree {i}:\n{tree}")

This provides details such as:

• nodeid
• depth
• split (which feature was used)
• split_condition (the threshold value)
• yes, no, missing, children
• leaf (leaf node weight)

However, I would like to access additional debugging information, specifically:

• Histogram bins used for each feature during training
• Candidate split points considered at each node
• Gain values for each candidate split

I understand that this information is not available through booster.get_dump(), and I suspect that modifying the C++ source code might be necessary. However, I am unsure about the exact location in the code where this information is processed and how to print it effectively.

What I have tried

• Setting verbosity=3, but this does not expose histogram details.
• Looking into src/tree/updater_hist.cc, where functions like BuildHist and EvaluateSplits seem to handle histogram construction and split evaluation.
• Considering adding LOG(INFO) or std::cout statements in the source code to debug this information.

Questions

1. Is there an existing way to enable detailed logging for histogram bins, split candidates, and gain values?
2. If modifying the C++ source is necessary, where would be the best place to insert logging statements to obtain this information?

Any guidance would be greatly appreciated! Thank you. 🙏