本地生活课程语音转文本大模型语料库数据

本数据通过语音大模型将音频课程或视频课程转为文本，使得大语言模型可以从音频数据中间接学习到知识，扩充模型可用的数据类型，提升模型效果。本数据作为自然语言处理模型训练的原材料，可用于各AI大模型学习和理解结构化数据，帮助AI大模型优化、校准、迭代升级，具有很强的复用性，适用于市面上所有大语言模型的训练。本数据中各个课程的点赞数、评论数等可以帮助使用者判断观众对各类型课程的接受情况，点赞、评论等数字大，代表观众对该类型课程更加喜爱，接受程度高，为课程创作提供方向。1、数据收集：饿了么域内存在大量的视频及音频数据文件，包含商家课程，电销通话等，对于大语言模型来说是非常丰富的学习资料，需要将多模态数据转化为文本格式，以供大语言模型学习，本数据便是将商家课程的视频文件转为文本的结果，其中content字段是转译后的结果。
2、数据处理：通过语音大模型，将工程存储在oss上的视频文件，转化为音频的文本文件，记录课程名称、课程介绍、观众数、点赞数、评论数、观看次数、课程风格、详细介绍、图像信息、内容、是否敏感、敏感词、内容长度、类别，经过文本标准化，敏感词过滤后可作为数据资产可用于下游大模型的预训练和RAG，course_style字段为课程风格，1代表图文课程，2代表视频课程，内容长度为“内容”字段所包含的字符总数。
3、对于C端等一些对数据质量要求较高的场景，评判数据集是否包含黄暴恐信息，通过语言识别模型将既定的敏感词与content字段进行过滤，输出是否敏感和敏感词两列结果，是否敏感字段结果为TRUE，代表有敏感词，对应显示敏感词内容，FALSE为没有敏感词，对应敏感词字段为无。

This dataset converts audio or video courses into text via speech LLMs, enabling large language models (LLMs) to indirectly acquire knowledge from audio data, expanding the available data types for models and improving model performance. As training raw material for natural language processing (NLP) models, this dataset can be used by various AI LLMs to learn and understand structured data, assisting in the optimization, calibration, and iterative upgrading of AI LLMs. It boasts strong reusability and is applicable to the training of all LLMs on the market. Metrics including the number of likes and comments for each course in this dataset allow users to assess audience acceptance of different course types. Higher values of likes and comments indicate that audiences prefer the corresponding course type more and have a higher acceptance level, providing guidance for course creation.

1. Data Collection: There are a large number of video and audio data files within the Ele.me domain, including merchant courses, telemarketing calls, etc., which serve as extremely rich learning resources for LLMs. It is necessary to convert multimodal data into text format for LLMs to learn from. This dataset is the result of converting video files of merchant courses into text, where the "content" field is the post-translation result.

2. Data Processing: Using speech LLMs, video files stored on Object Storage Service (OSS) are converted into audio-derived text files. The dataset records course name, course introduction, audience count, number of likes, number of comments, view count, course style, detailed introduction, image information, content, sensitivity flag, sensitive words, content length, and category. After text standardization and sensitive word filtering, it can be used as data assets for downstream LLM pre-training and Retrieval-Augmented Generation (RAG). The "course_style" field indicates the course type: 1 represents graphic-text courses, 2 represents video courses. The content length refers to the total number of characters contained in the "content" field.

3. Quality Filtering: For scenarios like consumer-facing (C-end) applications with high data quality requirements, this dataset is screened for pornographic, violent, and terrorist information. A language recognition model is used to match preset sensitive words against the "content" field, generating two output columns: "is_sensitive" and "sensitive_words". If the value of "is_sensitive" is TRUE, it indicates the presence of sensitive words, and the corresponding "sensitive_words" field displays the detected sensitive content. If the value is FALSE, it means no sensitive words are present, and the "sensitive_words" field is marked as "none".

搜集汇总

背景与挑战

背景概述

本地生活课程语音转文本大模型语料库数据是一个包含1002条记录的企业数据集，每日更新，用于自然语言处理模型的训练和优化。数据来源于饿了么域内的视频及音频文件，通过语音大模型转换为文本，包含课程名称、介绍、观众数等字段，适用于AI大模型的训练和课程创作方向的分析。