[MohammadWasi]

Summary

• make Comet ML an optional dependency for the Qlib fine-tuning workflow
• disable Comet logging by default because the checked-in configuration contains placeholder credentials
• keep the existing Comet behavior unchanged when use_comet is explicitly enabled

Problem

Both finetune/train_tokenizer.py and finetune/train_predictor.py imported comet_ml unconditionally. As a result, a user who follows the documented no-Comet path (use_comet = False) still fails before training starts unless Comet is installed.

Implementation

• add finetune/utils/experiment_logging.py to lazily create a Comet experiment only when enabled
• provide an actionable error when logging is enabled but comet_ml is missing
• reuse the helper from both fine-tuning entry points
• document how to opt in to Comet logging

Tests

• python -m unittest tests/test_experiment_logging.py
• python -m py_compile finetune/config.py finetune/train_tokenizer.py finetune/train_predictor.py finetune/utils/experiment_logging.py
• git diff --check

The new tests verify that disabled logging does not import Comet, enabled logging explains a missing dependency, and valid Comet configuration creates and configures the experiment.

[MohammadWasi]

Hi @AnMakc @ehan1990 @alexliao @xxsunyxx Can you please review this and share your feedacks?

[itsadrianxv]

Thanks — nice usability improvement. Moving import comet_ml out of module scope into utils/experiment_logging.py means fine-tuning runs without Comet installed, and defaulting use_comet = False avoids surprising failures for new users. Lazy import is the right pattern here, and the new unit test covers the helper. LGTM 🚀