How xming521/weclone works
Project Type and Purpose
• This is a Python-based one-stop solution (“WeClone”) for creating a digital avatar from chat logs.
• It combines a CLI frontend (weclone-cli), a data-processing pipeline, model fine-tuning (SFT/LoRA/pt), inference services (vllm), a simple REST API, a browser demo, and an audio cloning submodule.
Main Components
A. CLI Entrypoint
• weclone/cli.py – defines commands: make-dataset, train-sft, train-pt, export-model, webchat-demo, server, test-model.
B. Data Pipeline
• Raw chat CSVs under /dataset/csv → cleaning & filtering (weclone/data/clean/…, strategies.py) → QA-pair generation (weclone/data/qa_generator.py).
C. Training Pipeline
• SFT training (weclone/train/train_sft.py)
• PT training (weclone/train/train_pt.py)
• Model export (weclone/train/export_model.py)
• Relies on external heavy-compute libs: PyTorch, vllm, deepspeed, flash-attn.
D. Inference Service
• vLLM-based infer module (weclone/core/inference/vllm_infer.py)
• CLI demo (weclone/eval/cli_demo.py) & web demo (weclone/eval/web_demo.py) call this.
E. REST API
• FastAPI-style or Flask in weclone/server/api_service.py
• Serves endpoints for inference, health checks.
F. Browser-based Demo
• Static/HTML + JS in weclone/eval/web_demo.py, talks to local API.
G. Audio Cloning Submodule
• Directory weclone-audio/: TTS inference (Llasa, Spark-TTS), a minimal server under server未完工/.
H. Configuration & Artifacts
• settings.jsonc – tuning, model paths, LoRA parameters.
• ds_config.json – data-split config.
• Local FS stores datasets, intermediate JSON, model checkpoints, logs.
I. External Dependencies
• LLM provider (local Qwen2.5/BLOOM, etc.)
• AstrBot (as a downstream consumer via OpenAI-compatible API).
Relationships and Data Flows
• CLI → invokes Data Pipeline → writes cleaned dataset.
• CLI → invokes Training Pipeline → reads cleaned dataset, calls external LLM frameworks → writes model artifacts.
• CLI/server → REST API loads model artifacts → on request calls Inference Service → returns JSON responses.
• Browser Demo ↔ REST API (HTTP).
• AstrBot or any OpenAI-compatible client ↔ REST API.
• Audio submodule server ↔ Inference Service (text output) → generates audio → returns WAV/stream.
Architectural Patterns & Principles
• Modular layer separation:
1. Presentation (CLI, Web Demo)
2. Application (Data, Train, Inference, API)
3. Infrastructure (File System, External LLM libs, GPU resources)
• CLI orchestrator pattern: single binary dispatches to submodules.
• Micro-services style for REST API and audio server (can be containerized).
• Config-driven behaviour via JSONC templates.
Essential Elements
• Details: note files or modules under each component (e.g., weclone/data/clean/get_score.py resides in Data Pipeline).