In the wave of live commerce, the Streamer-Sales large model brings revolutionary changes to live sales with its unique intelligent features. This is not just a tool; it is an intelligent companion for live commerce, capable of automatically generating engaging commentary scripts, answering customer questions in real time, and making sales more vivid and efficient.
Video by project author HinGwenWoong
High-definition video address:
https://www.bilibili.com/video/BV1ZJ4m1w75P/?spm_id_from=333.337.search-card.all.click&vd_source=1f648bd6809721ebf81314c53bda24f8
Key Features:
One-click host script generation: Streamer-Sales can automatically generate commentary scripts based on product features, stimulating users' purchase desires and enhancing sales conversion rates.
Inference acceleration: Integrated with LMDeploy technology, Streamer-Sales has significantly improved inference speed, supporting KV caching and Turbomind, making live interactions smoother.
Retrieval-Augmented Generation (RAG): Combining product manuals and related documents, Streamer-Sales can generate more practical script content.
Automatic Speech Recognition (ASR): Converts the host's speech into text in real time, facilitating interaction with the audience.
Text-to-Speech (TTS): Emotional voice output makes product commentary more vivid and natural.
Digital human generation: Through virtual host videos, Streamer-Sales provides more vivid product commentary, enhancing audience experience.
Real-time information query (Agent): Capable of real-time query of information such as delivery status, providing users with the latest data.
Model Architecture:
Streamer-Sales is a model fine-tuned based on InternLM2, integrating multiple advanced technologies to build a comprehensive live commerce system. Its architecture includes data generation and processing, model training and fine-tuning, inference and generation, voice processing, digital human generation, and real-time information query.
Model Introduction:
streamer-sales-lelemiao-7b: A model specifically designed for generating commentary scripts for live commerce, trained with a large amount of product data to produce high-quality scripts.
Model download address: https://top.aibase.com/tool/maihuozhubodamoxing-lelemiao-7b
streamer-sales-lelemiao-7b-4bit: The 4bit quantized version of the model, optimized for inference speed and resource usage, suitable for resource-constrained environments.
Model download address: https://top.aibase.com/tool/maihuozhubodamoxing-lelemiao-7b-4bit
Applicable Scenarios:
Online live sales, offline store promotions, product ad script generation, Streamer-Sales excels in all these areas.
Technical Highlights:
Dataset generation: Uses Tongyi Qianwen and Wenxin Yiyan to generate datasets, including product scripts and Q&A dialogues.
Model fine-tuning: Uses xtuner to fine-tune the base model to adapt to different products and user needs.
Quantization processing: 4bit quantization improves inference efficiency while maintaining generation quality.
Currently, the Streamer-Sales project has been open-sourced on GitHub, providing model download links and an online experience platform, allowing live commerce hosts to easily integrate this powerful AI assistant and enjoy its convenience and advantages.
Project address: https://top.aibase.com/tool/streamer-sales
Online experience: https://openxlab.org.cn/apps/detail/HinGwenWong/Streamer-Sales