Storage for file system.

PromptedVQAGenerator reads conversation history with optional image/video inputs to generate answers. Supports both multimodal mode (with ima

Client for interacting with a Vision-Language Model (VLM) via OpenAI's API. Provides methods for single-image chat, batch image processing,

A class for generating text using vllm, with model from huggingface or local directory

[Refine] 通用 Python 逻辑精炼器。 用于：分句、正则解析、字符串拼接等不涉及模型调用的数据加工。

通用多占位符 Prompt 模板类。 模板示例： template = ( "Descriptions:\n" "{descriptions}\n\n" "Collect all detail

The prompt for custom code operations.

CARDINAL类

Subclass torch's LayerNorm to handle fp16.

PromptTemplatedQAGenerator: 1) 从 DataFrame 读取若干字段（由 input_keys 指定） 2) 使用 prompt_template.build_prompt(...) 生成纯文本 prompt 3) 将该 prompt

The prompt for the video quality evaluator.

Generate QA conversations from video captions via a prompt template. This operator rewrites the first user message in each conversation to a

DataFlow operator: consume video path/info/scene columns and produce per-row clip lists. Output column is configurable (default: 'video_clips

DataFlow operator: - 从 storage 的 dataframe 中取 `video_clips_key` 列（list 的结构） - 展平并仅切割 filtered == False 的片段 - 将成功结果 DataFrame 写入 `csv_

DataFlow operator: scan dataframe for video files, probe basic info, and write the processed dataframe back to storage.

DataFlow operator: detect scenes for each video path from the input dataframe, and write the processed dataframe (adds 'video_scene' column)

中文数字/数位字符 每个字符除繁简体外还有一个额外的大写字符 e.g. '陆' 和 '陸'

Generate captions from videos by prompting a VLM service. This operator rewrites the first user message in each conversation with a template-

[Refine] 视觉一致性精炼器。 输入：文本列表 (sentences/answers) + 图片。 行为：对列表中每一项进行视觉验证 (Yes/No)，去除幻觉内容。

[Generate] 批量视觉问答生成器。 输入：问题列表 (Questions) + 图片。 输出：答案列表 (New Text Content)。

FixPromptedVQAGenerator generate answers for questions based on provided context. The context can be image/video.

Evaluate text-based answers across multiple question types. This evaluator supports multiple question types with appropriate scoring metrics:

PromptTemplatedVQAGenerator: 1) 从 DataFrame 读取若干字段（由 input_keys 指定） 2) 使用 prompt_template.build_prompt(...) 生成纯文本 prompt 3) 将该 prompt

通用分数过滤算子：根据指定列的分数值进行过滤

[Generate] 使用通用 VLM Serving 生成 BBox 数据。 输入：Image + Keywords List 输出：BBox Map

The prompt for the video Chain-of-Thought QA generator.

Container for a single video's probed metadata.

[Refine] 视觉依赖性校验器。 对 MCQ 列表进行“旋转 + 双盲测试 (Visual vs Text-Only)”： 筛选出 必须依赖视觉信息 (High Visual Acc) 且 不能仅凭常识 (Low Text Acc) 的问题。

[Generate] 调用 VLM 生成推理链。 支持 Fallback：如果 input_existing_chains_key 中已有数据，则直接使用，不进行生成。

文本格式规范化 + WikiQA 格式解析算子。

CTCForcedAlignmentSampleEvaluator is a class that performs CTC forced alignment on audio data.

Sample representing a single clip to be processed.

DIGIT类

Caption Generator is a class that generates captions for given images.

Prompt definitions for CapRL/VisualOnlyMCQ pipeline.

A class for multimodal generation using sglang Engine, 支持从 HuggingFace 或本地目录加载模型。

[Refine] 解析 MCTS 树结构，提取推理链。

PromptedVQA is a class that generates answers for questions based on provided context.

The registry that provides name -> object mapping, to support third-party users' custom modules. To create a registry (e.g. a backbone r

A single processed scene segment.

封装了一个独立的 Silero VAD 模型实例。 负责加载模型、读取音频，并执行VAD算法。

TELEPHONE类

DataFlow operator: compute aesthetic scores and filter clips based on threshold. This operator combines VideoAestheticEvaluator's scorin

DataFlow operator: extract frames for each clip from the upstream dataframe. Consumes `video`, `video_info`, `video_clips`; produces `video_f

DataFlow operator: compute luminance statistics and filter clips based on thresholds. This operator combines VideoLuminanceEvaluator's a

DataFlow operator: generate merged caption structures from multiple clips. Input: DataFrame with individual clip captions Output: Up

DataFlow operator: compute OCR scores and filter clips based on thresholds. This operator combines VideoOCREvaluator's analysis function

CLIP image encoder + MLP head.

通用的批处理 Wrapper。 在静态检查/IDE 里，BatchWrapper.run 的签名和 operator.run 完全一致。 运行时，会把 operator.run 的 __doc__ 和 __signature__ 也拷过来， 这样 help(bw.

Evaluate CLIP scores between images and text, adding scores to dataframe. Computes the cosine similarity between CLIP embeddings of imag

The prompt for the AutoPromptGenerator.

中文数字字符

使用 CLIPEvaluator 对 image + caption 进行图文对齐打分。

Dataset that loads up to `load_num` frames per clip. Missing/short frame sets are padded by repeating the last frame.

Dataset that loads up to `load_num` frames per clip for OCR processing.

Dataset that loads up to `load_num` frames per clip. Missing/short frame sets are padded by repeating the last frame; If no frame is foun

一行命令即可完成图片批量 ContextVQA Caption 生成。

一行命令即可完成图片批量 ContextVQA Caption 生成。

DATE类

Filter operator that removes successful responses and keeps only failed ones. Uses a rollout model to generate responses and a verifier

FRACTION类

一行命令即可完成图片批量 Caption 生成。

一行命令即可完成图片批量 Caption 生成。