Self-AI: Practice of LLM Self-Evolution Fine-Tuning Framework Based on Unsloth and LoRA

self-AI (project code: Neuro_Live) is an open-source experimental framework focused on LLM fine-tuning and self-optimization. Its core goal is to achieve efficient training and inference on consumer-grade hardware. The framework integrates Unsloth acceleration technology and LoRA parameter-efficient fine-tuning method, supports automated data evolution cycles and 4-bit quantized inference, providing a feasible solution for low-cost LLM personalized customization.

Against the backdrop of widespread LLM applications, low-cost implementation of model personalized customization is a focus for developers. As an open-source experimental framework, self-AI integrates cutting-edge LLM engineering technologies, providing a complete experimental platform for researchers and tech enthusiasts, aiming to lower the technical threshold for LLM training and inference on consumer-grade hardware.

1. Unsloth Acceleration Engine: Optimizes attention mechanisms and gradient computation, increasing training speed by over 2x compared to native Transformers, reducing memory consumption by 60%, and supporting consumer-grade GPUs; 2. LoRA Fine-Tuning: Trains only low-rank adaptation matrices to reduce resource consumption, supports multi-task adapter switching, and merge_model.py enables dynamic merging of models and adapters; 3. 4-bit Quantized Inference: Reduces weight precision, achieves an inference speed of over 25 tokens per second, and adapts to resource-constrained environments.

Automated data evolution is implemented via the evolve_neuro.py module: extract samples from user interactions → archive to data/history_growth.jsonl for long-term accumulation → trigger incremental training → form an "Interaction-Learning-Evolution" closed loop. Hierarchical storage design: data/growth_data.jsonl stores temporary memory, history_growth.jsonl stores historical evolution data, balancing real-time performance and persistence.

Modular design: The src directory contains core modules (core is responsible for inference and embedding similarity memory retrieval), interaction modules (body handles VTS body expression control), and script modules (scripts provide entry points for training/dialogue/startup/evolution, etc.). Multimodal extension: Reserved TTS interfaces, ref_audio directory stores reference audio, and supports custom environment variables to specify paths.

Environment requirements: CUDA 12.1+ NVIDIA GPU, Python needs to install dependencies such as PyTorch and Unsloth. Usage process: 1. Environment preparation (install dependencies); 2. Model training (run train.py); 3. Interaction test (chat_neuro_v2.py); 4. Start service (start_neuro.py). The project is a lightweight repository, and users need to prepare model weights by themselves.

Application scenarios: Personalized AI assistant customization, domain-specific knowledge injection, virtual anchor development, LLM training technology learning and research. Current limitations: Self-evolution depends on the quality of interaction data, lacks data cleaning; multimodal only has reserved interfaces; long-term memory management strategy is not clear. Improvement directions: Introduce RLHF to improve self-evolution quality, intelligent memory management, improve multimodal capabilities, and visual monitoring interface.

self-AI demonstrates a feasible path for LLM engineering applications: reducing training and deployment costs via Unsloth and LoRA, enabling continuous learning through self-evolution mechanisms, and supporting expansion with modular design. It provides a valuable reference implementation for developers exploring LLM personalized customization on consumer-grade hardware.