Remove parent directory name.

Avoid bugs in mps backend by not using in-place operations.

Save current rank's data to temporary file

Save samples as HuggingFace Dataset

Save samples as sharded Parquet files

Collect task, task_id, and hidden_state for each sample into a global list

Collect task, task_id, plan, hidden_state, task_type, and task_level into a global list

加载训练好的模型和 tokenizer（plan-only 模型）。

统一加载： - HF 导出目录（含 hidden_mha_state.pt） - Trainer checkpoint（checkpoint-XXXX/，pytorch_model.bin 是 wrapper 的完整 state_dict） - 仅基础模

Initialize distributed training environment

Initialize distributed training environment

Initialize distributed training environment with robust error handling

Stream output from a role.

Stream output.

Convert config to dictionary

Validate and process arguments

Validate and process arguments

messages: [{"role": "user"/"assistant", "content": str}, ...] plan: 纯文本的 step-by-step plan，会被插在第一条 user 输入的 '<bop>' 和 '<eop>' 之间

The history should be a format like: [ {"role": "system", "content": "You are a helpful assistant."}, {"role"

Initialize the evaluator Args: model_name_or_path: Path to model or HuggingFace model name tokenizer_name: T

model_reference: the actual model being trained (may be wrapped by DDP/FSDP; needs unwrapping)

Build a new LiteLLM model object that represents a model in a given state.