Get available memory for each GPU.

Execute a three-layer reasoning chain with different roles and prompts

Execute a reasoning chain and save the sample (including task_type / task_level)

将plan文本插入到指定位置，前后加入特殊标记

Check whether the output is a complete sentence.

List all available predefined configurations Returns: Dictionary mapping config names to descriptions

从 DeepSpeed 的 mp_rank_00_model_states.pt 中恢复 - model.adaptive_proj.* 参数 - model.scale / model.output_scale（如果在 ckpt 里有） 同时： - 优先

Load dataset configurations

Load the dataset for the specific task

Load model and prepare for DDP

Load model and prepare for DDP

Load model and tokenizer with error handling

Load model configurations

Main evaluation script

Main execution function

Main execution function

can be overrided to costomize the conversation template for different model workers.

Create dataset and collator for supervised fine-tuning.

Build data module.

Merge data from all ranks (called by rank 0)

将hidden_state的分布调整为与reference_embeds相似

Check whether the text violates OpenAI moderation API.

Parse command line arguments

Parse command line arguments

Preprocess conversations and track the position of the first human message end marker.

Print a semaphore in better format.

Print output.

Process a batch of task items and collect data

Prompt for input from a role.

RandomRot: H' = μ̂ + (H - μ̂) Σ̂^{-1/2} Q Σ̂^{1/2} Where Q is a Haar random orthogonal matrix, ensuring: - Exactly preserve mean / covari