r"""Freeze vision tower and language model for VLM full/freeze tuning.

Prefer the *dataset-assigned* bucket if present, so that histograms match your final dataset quotas after adjacent fills. We intentionall

r"""Get a list of scores of the reward model.

r"""Load model config.

把 'B' / '(B)' / '选项B' / 'option:B' / 'Choice b.' 等归一为 'B'；否则返回 None。

r"""Remove thought from assistant message.

支持 .gz 与普通文件

r"""Add or replace eos token to the tokenizer.

r"""Build a dictionary containing the current training configuration.

参考你的“提问方式”： - system: You are a helpful assistant. - user: "Please reason step by step, and put your final answer within \\bo

r"""Check if the video is nested images.

r"""Remove args with NoneType or False or empty string value.

r"""Clean up the server process when the engine is destroyed.

r"""Convert elements to token ids.

r"""Get the path to user config.

r"""Return the jinja template.

r"""Return the merged datasets in the standard format.

r"""Compute video sample indices according to fps.

r"""Start the training process.

r"""Build and validate the evaluation arguments.

r"""Parse eval example. input: a dict with keys {"question", "A", "B", "C", "D", "answer"} output: a tuple of (prompt, response).

r"""Build and validate the training arguments.

r"""Pre-process a single image.

r"""Preview the training commands.

仅在最外层（()[]{}) 深度为 0 时按逗号切分；若整体被 () 或 [] 包着，认为不是多解列表。