// the take
Streamer-Sales is a fine-tuned InternLM2-7B model trained to play the role of a Chinese live-stream shopping host — it generates sales scripts, answers buyer questions in character, and supports RAG against product manuals. The full system layers on TTS (GPT-SoVITS), ASR (FunASR), a talking-head digital human (MuseTalk), and a FastAPI/Vue admin backend. It's aimed at Chinese-market e-commerce teams or researchers studying conversational sales AI.
The data pipeline is unusually well-documented: the YAML-driven dataset generator, the OCR-to-manual pipeline, and the published training JSONL with ~1500 conversations let you reproduce or extend the fine-tune without reverse-engineering anything. LMDeploy TurboMind integration is a real win — the benchmark shows 5x throughput improvement with 4-bit quantization, dropping from 61 to 307 words/s, which matters a lot for real-time streaming. The microservice decomposition (LLM, TTS, ASR, digital human as separate FastAPI services behind one middleware) is sensible for production: you can scale or swap each independently. RAG via product manuals means you can onboard new products without retraining — just upload a manual and rebuild the vector store.
The whole stack is China-specific and largely undocumented in English, making it a hard adopt for anyone outside that context — the model, the training data, the TTS voices, and the Agent APIs (delivery tracking, weather) are all tuned for Chinese. The digital human generation isn't real-time: it uses ComfyUI workflows and MuseTalk to produce video clips offline, not a live avatar, which is a significant gap versus the README's framing. The only supported base model is InternLM2-chat-7b, and the fine-tuning configs are hardcoded to it — switching to Qwen2 or LLaMA3 requires non-trivial work with no guidance. Docker Compose support is backend-only (the README says so in two places), meaning the Vue frontend still requires manual Node setup, which undermines the 'one-click deploy' claim.
We read the README, the directory tree and a key source file in June 2026. Nobody pays to be listed. How we judge →