Clear model and release memory.

Compute log probabilities of given sequences.

Compute log probabilities of given sequences.

Compute log probabilities of given sequences.

Compute log probabilities of given sequences.

Decode base64 string to audio. Args: base64_string: Base64 encoded audio string Returns: tuple: (sa

Decode base64 string to video frames. Args: base64_string: Base64 encoded video string Returns: Lis

Open an image from various sources (base64, URL, local path) and encode it as base64. Args: image_path: Path or data of the imag

Open an image from various sources (base64, URL, local path) and encode it as base64. Args: image_path: Path or data of the imag

Open an image from various sources (base64, URL, local path) and encode it as base64. Args: image_path: Path or data of the imag

Extract preference (ResponseA or ResponseB) and reasons for each category from model output Args: output: Model output text

Forward pass through the RMSNorm layer. Args: x (torch.Tensor): The input tensor. Returns: torch.Te

Forward pass through the RMSNorm layer. Args: x (torch.Tensor): The input tensor. Returns: torch.Te

Load the houses from a .jsonl.gz file.

Compute advantages and returns using Generalized Advantage Estimation (GAE).

Get parameter groups with customized weight decay value.

Get reference answers from input data Args: input_data (list[InferenceInput]): input data inference_outputs (list[Infe

Computes the shortest path to an object from an initial position using a controller :param object_id: string with id of the object

Computes the shortest path to an object from an initial position using a controller :param controller: agent controller :para

Given an input task spec, returns a list of all synsets relevant to that task's success.

GPT-based alignment evaluation function for assessing alignment between internal beliefs and external outputs. Args: inner_belie

GPT-based consistency evaluation function for assessing consistency between CoT and final output. Args: cots: List of chain of t

GPT-based CoT and internal belief alignment evaluation function. Args: cots: List of chain of thoughts inner_beliefs: Li

替代 litellm.batch_completion 的 HuggingFace 本地推理函数。 Args: model_path: 本地模型路径或 HuggingFace repo 名称。 messages: 一个列表，列表中的每一项是一个 c

Initial configuration checking.

Check if the response provides helpful technical information

Check if the response provides helpful technical information

Loss function for supervised finetuning.

e.g. in the bedroom

e.g. in the bedroom

Model inference. Support multi-round inference. Args: input_data (list[InferenceInput]): input data Returns: