Lokum-F is an advanced, standalone desktop application built for Apple Silicon (M-Series) Macs. It provides a comprehensive GUI for local Large Language Model (LLM) fine-tuning, Retrieval-Augmented Generation (RAG) indexing, and model evaluation—all leveraging the native power of the Apple MLX framework.
- Local LoRA Fine-Tuning (MLX): Effortlessly train and adapt large language models entirely on-device using Apple's MLX architecture. Configure advanced parameters (Rank, Alpha, Batch Size, Layers) via an intuitive UI without writing any CLI commands.
- Intelligent Adapter Management: Automatically tracks and organizes fine-tuned adapters with deterministic hashing (
R16_A32_B1_L16_A7B2C) for pristine developer experience (DX). - One-Click Model Fusion: Merge your trained LoRA adapters into base models seamlessly with a single click. Includes auto-cleanup to delete heavy adapter artifacts post-fusion, saving valuable disk space.
- Smart RAG Engine (FAISS): Build robust contextual memories for your models. The engine supports dynamic re-indexing, chunk compaction, and active garbage collection to prevent index bloat and ensure fast retrieval.
- Developer-Friendly Diagnostics: Real-time MLX log parsing that filters out Apple Metal API spam ("God-Mode" limits) and provides actionable, clear hints for common hardware errors (e.g., Layer mismatch, Out-Of-Memory exceptions).
- Speech-to-Text Integration: Built-in
mlx-whisperintegration for instant audio transcription directly within the chat interface.
- Apple Silicon Native: Fully optimized for M-series unified memory architecture.
- PyQt6 GUI: A modern, responsive desktop interface with built-in presets (Ultra, Good, Mid, Low, Custom) scaling to your available VRAM.
- Vector Database: FAISS-based fast semantic search using optimized embeddings.
- Create a virtual environment:
python3 -m venv .venv
- Activate the environment:
source .venv/bin/activate - Install dependencies:
pip install -r requirements.txt mlx-whisper sounddevice scipy numpy
- Run the application:
python3 main.py
When using the "Custom" preset, remember the Golden Rule for Apple Silicon memory stability:
- Rank (r): Defines the learning capacity. Keep it an even power of 2 (8, 16, 32).
- Alpha: The scaling factor. Should be
2 * Rankor at least equal to Rank. - Train Layers: Do not exceed your model's physical layer count. Reducing this (e.g., to 16 or 8) drastically reduces VRAM consumption.
Built for local AI development on macOS.