[Patch] Skip calculating loss of empty <think></think> (ಥ﹏ಥ)

[HAOCHENYE]

No description provided.

[Copilot]

Pull request overview

This PR introduces a temporary monkey patch to skip loss calculation for empty <think></think> tags in training data. The patch is controlled via the XTUNER_SKIP_EMPTY_THINK environment variable and is explicitly designed to be disposable rather than a permanent solution.

Key Changes:

• Added _hardcode_patch.py with _SkipEmptyThink mixin class that filters empty thinking tags from labels
• Modified multiple tokenize function classes through runtime base class injection
• Wrapped tokenization methods to apply label processing post-tokenization

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 7 comments.

File	Description
xtuner/v1/datasets/_hardcode_patch.py	Implements the core monkey patching logic with _SkipEmptyThink mixin and label processing wrapper for filtering <think></think> sequences
xtuner/v1/datasets/__init__.py	Imports the hardcode patch module to activate the monkey patching when the environment variable is set

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Typo in comment: "faild" should be "fail".

Suggested change

	# Need to inherit ABC otherwise the assignment of `__bases__` will faild
	# Need to inherit ABC otherwise the assignment of `__bases__` will fail

[Copilot]

The __init__ method accepts **kwargs but never uses or forwards them. This could silently ignore important configuration parameters that were intended for the parent class or other mixins in the MRO chain. Consider using super().__init__(tokenizer=tokenizer, **kwargs) to properly forward any additional arguments.

Suggested change

	super().__init__(tokenizer=tokenizer)
	super().__init__(tokenizer=tokenizer, **kwargs)