Send a chat completion request to the OpenAI API and return the generated content. :param model: Model name for the request.

Convert timecode to seconds. Accepts integer seconds or 'HH:MM:SS.mmm' string format.

Parse 'HH:MM:SS.mmm' -> seconds as float; return None if invalid.

Validate the model and API configuration.

Apply aesthetic score filtering to clips. Args: dataframe: DataFrame with video_clips column containing aesthetic_score

Apply luminance filtering to clips. Args: dataframe: DataFrame with video_clips column containing luminance_mean video_c

Apply OCR score filtering to clips. Args: dataframe: DataFrame with video_clips column containing ocr_score video_clips_

输入: messages -> list of list (每个元素是聊天序列) 输出: full_prompts -> list of dict {prompt, multi_modal_data}

Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode strings. This means you

清洗字符串中的 markdown 标记，比如 **bold** 或 *italic*。 用于处理嵌套 bold (e.g. **Title **Sub** End**) 的情况。

根据数字系统类型返回创建相应的数字系统，默认为 mid NUMBERING_TYPES = ['low', 'mid', 'high']: 中文数字系统类型 low: '兆' = '亿' * '十' = $10^{9}$, '京' = '兆' * '十', et

r""" Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded prompt_2 (`str` or `List[s

Instantiate the MLP head if not yet created; optionally load weights.

Extract content between <answer> and </answer> tags.

提取大模型回复中的选项字母

Get the object map of the registry.

Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings).

处理单个音频文件的完整流程。

类似于 vision 的 process_vision_info，从 message 列表中提取音频输入。 支持三种格式输入： - 本地或 http(s) URL 路径（通过 librosa 接口处理） - base64 编码 (data:audio/…;b

Compute aesthetic scores per clip and inject them back into `video_clips`. Args: dataframe: upstream df with `video` and `video_clip

Compute luminance stats (mean/min/max) per clip and inject them into `video_clips`. Args: dataframe: upstream df with `video` and `v

Compute OCR scores per clip and inject them into `video_clips`. Uses DataLoader for efficient batch processing with multi-worker data loading

Convert seconds to integer frame index using fps.

Convert seconds (float) to 'HH:MM:SS.mmm' timecode.

混淆选项逻辑 (保持原版业务逻辑不变)

Split a string into paragraphs.

Convert 'HH:MM:SS.mmm' timecode into seconds (float).

把一个 OperatorABC 实例包装成按 batch_size 批处理的 Wrapper 同时让 wrapper.run 的函数签名 = operator.run 的签名

上下文管理器，确保视频资源正确释放

动态加载类。 :param item: 类名 :return: 动态加载的类对象

Args: name (str): the name of this registry

Apply all text filters. Returns True if passed, False if failed.

Compute width/height safely.

构建 OpenAI API 格式的消息内容 :param prompt: 文本提示 :param image_paths: 图像路径列表 :param video_paths: 视频路径列表 :par

从对话历史构建 OpenAI API 格式的消息列表 :param conversation: 单个对话历史，格式为 [{"role": "user", "content": "..."}, ...] :param image_pa

Build OpenAI image editing request (supports DALL-E 2 and gpt-image-1) Returns: (files_dict, data_dict): files dictionary

Build a single OpenAI image generation request (supports DALL-E 2/3 and gpt-image-1)

Build OpenAI image variation request (only supports DALL-E 2) Returns: (files_dict, data_dict): files dictionary (for mul

Build one prompt per row using the configured template. Unlike the QA variant, this template does not depend on row fields, so we sim

Build one prompt per row using the template and the caption column. Raises: KeyError: if the required caption column is missing.

Build one prompt per row using the template and problem information.

Build one prompt per row using the template and problem information.

Build one prompt per row using the template and problem information.

Build reformatting prompts for each video with valid parsed reasoning. Args: df: DataFrame containing parsed_rea

Build a list of ClipSample objects from the dataframe. Similar to video_aesthetic_evaluator._build_samples_from_df.

Build samples for whole-video aesthetic scoring (no clip splitting required). Each row produces a single sample, frames will be sampled from

Build samples for whole-video OCR (no clip splitting required). Each row produces a single sample with frames sampled directly from the video

Build samples for whole-video luminance scoring (no clip splitting required). Each row produces a single sample; frames are sampled from vide

Calculate informativeness using singular value entropy (SVE) of multi-modal token embeddings. Based on DataTailor paper equa

Calculate representativeness using inter-cluster similarities. Based on DataTailor paper equations (5) and (6): τ_i^

计算调整后的尺寸

Calculate uniqueness within clusters using weighted distances. Based on DataTailor paper equation (4): V_i^Uni = ∑||

Calculate uniqueness scores based on k-nearest neighbor similarity. Args: embeddings: Image embeddings of shape

Send API request

Call Gemini API to generate images using google.genai library

Use chat mode for multi-round image editing Returns: list of images generated in the last round

Call OpenAI image editing API (/images/edits)

Call OpenAI image variation API (/images/variations)

Check basic image properties.

Check if image contains NSFW content. Returns True if safe, False if NSFW.

Perform hierarchical clustering on sample embeddings.

Compute CLIP scores between images and their corresponding texts. Args: images: List of images in various format

计算两帧之间的光流

Determine which local indices (0..num_frames-1) to save. - If interval_sec provided & fps>0 -> stride by round(interval_sec*fps), min 1.

Compute luminance statistics for a stack of frames. Args: images_uint8: (N, C, H, W) uint8 in [0,255], C=3 RGB Returns: (

Compute OCR text area ratio for a list of images. Args: images: List of image arrays (numpy arrays from cv2) clip_height

计算单个视频的运动分数

将格式1的数据转换为格式2。 支持多轮对话和多模态（图片、视频、音频），并验证token数量。