Training Parameter File Usage Guide

Parameter Creation

Parameters can be created through the frontend interface following the Training Parameter Profile specifications.

---

Parameter Usage in Training Code

Basic Parameter Handling

The Python template file custom-template.py pre-defines all optional/required parameters for training tasks. To use custom parameters from your configuration file, extend the existing argument parser while maintaining original parameter reception.

Example Configuration:

Parameter Label	Parameter Name	Value	Type	Required
Num Train Epochs	num_train_epoch	1	int	Yes

Implementation:

# Argument parsing
parser = argparse.ArgumentParser()
parser.add_argument('--num_train_epoch', type=int, required=True)

# Parse arguments
known_args, unknown_args = parser.parse_known_args()
args = known_args

# Parameter assignment
NumTrainEpoch = args.num_train_epoch

Critical Notes: 1. Type consistency between frontend definition and Python code is mandatory 2. Required parameters must have matching required=True declarations 3. Debug using argument inspection:

print("Known args:", known_args)    # Successfully parsed parameters
print("Unknown args:", unknown_args) # Unrecognized parameters

---

Optional Parameter Handling

For optional parameters, use conditional dictionary unpacking pattern:

trainer = SFTTrainer(
    args=TrainingArguments(
        # Optional parameter injection
        **({"num_train_epochs": NumTrainEpoch} if NumTrainEpoch is not None else {}),
        **({"max_steps": MAX_STEPS} if MAX_STEPS is not None else {}),
        # Other fixed parameters
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        ...
    ),
    ...
)

Best Practices:
1. Always provide fallback empty dict {} for conditional parameters
2. Maintain parameter naming consistency between configuration and code
3. Document optional parameters in both frontend labels and code comments