[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:mlx-lora-studio":3},"\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002FGoekdeniz-Guelmez\u002Fmlx-lora-studio\u002FHEAD\u002FSources\u002FMedia\u002Flogo_ultra-wide.png\" alt=\"MLX LoRA Studio\" width=\"100%\" \u002F>\n\u003C\u002Fp>\u003Cp align=\"center\">\n  \u003Cstrong>A native Mac App for LLM fine-tuning on Apple Silicon — fully on-device, fully open source.\u003C\u002Fstrong>\n\u003C\u002Fp>\u003Cp align=\"center\">\n  \u003Cem>\n    There is a quiet kind of power in running a model on the machine that sits in front of you.\u003Cbr \u002F>\n    MLX LoRA Studio puts fine-tuning on your Mac — local, native, and visible end to end.\u003Cbr \u002F>\n    Pick a model, choose an algorithm, watch the loss fall. The cloud is optional, the code is optional, the mystery is not.\n  \u003C\u002Fem>\n\u003C\u002Fp>\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FIMPORTANT-FIRST%20LAUNCH%20STEP-red?style=for-the-badge\" alt=\"Important first launch step\" \u002F>\n\u003C\u002Fp>\u003Cdiv>\n  \u003Cstrong>Important macOS first-launch step after installing:\u003C\u002Fstrong>\u003Cbr \u002F>\u003Cbr \u002F>\n  Because this first open-source release is distributed outside the Mac App Store and is not yet notarized, macOS may show a scary \"damaged\", \"cannot be opened\", or unidentified developer warning. This does \u003Cstrong>not\u003C\u002Fstrong> mean the app is malware. It is Apple's quarantine flag on a freshly downloaded app bundle.\u003Cbr \u002F>\u003Cbr \u002F>\n  After dragging \u003Cstrong>MLX LoRA Studio.app\u003C\u002Fstrong> into \u003Cstrong>\u002FApplications\u003C\u002Fstrong>, and before opening it for the first time, open Terminal and run:\n  \u003Cpre>\u003Ccode>sudo xattr -dr com.apple.quarantine \"\u002FApplications\u002FMLX LoRA Studio.app\"\u003C\u002Fcode>\u003C\u002Fpre>\n  Then open the app normally from \u003Cstrong>\u002FApplications\u003C\u002Fstrong>. You only need to do this once per installed copy.\n\u003C\u002Fdiv>\u003Chr \u002F>\n\u003Ch2>Table of contents\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"#release-v200\" rel=\"nofollow ugc noopener\">Release v2.0.0\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#why-mlx-lora-studio\" rel=\"nofollow ugc noopener\">Why MLX LoRA Studio?\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#what-it-is\" rel=\"nofollow ugc noopener\">What it is\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#what-it-isnt\" rel=\"nofollow ugc noopener\">What it isn't\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#features-at-a-glance\" rel=\"nofollow ugc noopener\">Features at a glance\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#screenshots\" rel=\"nofollow ugc noopener\">Screenshots\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#installation\" rel=\"nofollow ugc noopener\">Installation\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#quick-start-60-seconds\" rel=\"nofollow ugc noopener\">Quick start (60 seconds)\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#a-guided-tour-of-the-app\" rel=\"nofollow ugc noopener\">A guided tour of the app\u003C\u002Fa>\u003Cul>\n\u003Cli>\u003Ca href=\"#train\" rel=\"nofollow ugc noopener\">Train\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#live-metrics\" rel=\"nofollow ugc noopener\">Live Metrics\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#upload-to-hf\" rel=\"nofollow ugc noopener\">Upload to HF\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#algorithm-guide\" rel=\"nofollow ugc noopener\">Algorithm Guide\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#runs\" rel=\"nofollow ugc noopener\">Runs\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#settings--onboarding\" rel=\"nofollow ugc noopener\">Settings & Onboarding\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#the-training-pipeline-under-the-hood\" rel=\"nofollow ugc noopener\">The training pipeline under the hood\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#supported-training-methods\" rel=\"nofollow ugc noopener\">Supported training methods\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#configuration-reference\" rel=\"nofollow ugc noopener\">Configuration reference\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#memory--hardware-expectations\" rel=\"nofollow ugc noopener\">Memory & hardware expectations\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#building-from-source\" rel=\"nofollow ugc noopener\">Building from source\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#project-layout\" rel=\"nofollow ugc noopener\">Project layout\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#contributing\" rel=\"nofollow ugc noopener\">Contributing\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#star-history\" rel=\"nofollow ugc noopener\">Star History\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#license\" rel=\"nofollow ugc noopener\">License\u003C\u002Fa>\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"#acknowledgments\" rel=\"nofollow ugc noopener\">Acknowledgments\u003C\u002Fa>\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Chr \u002F>\n\u003Ch2>Release v2.0.0\u003C\u002Fh2>\n\u003Cp>\u003Cstrong>MLX LoRA Studio v2.0.0 updates the app for \u003Ccode>mlx-lm-lora\u003C\u002Fcode> 3.0.0.\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>This release adds FTPO, Dynamic Fine-Tuning, and memory-bounded Chunked NLL while aligning\nthe native training runner and documentation with the \u003Ccode>mlx-lm-lora\u003C\u002Fcode> 3.0.0 API. Synthetic\ndataset creation has been removed from the active app workflow, while historical synthetic\nrun folders remain readable in the Runs archive.\u003C\u002Fp>\n\u003Cp>If you need detialed and longer explanations for the algorythms used, then visit the \u003Ca href=\"https:\u002F\u002Fgoekdeniz-guelmez.github.io\u002FMLX-LoRA-Studio\u002F\" rel=\"nofollow ugc noopener\">wiki\u003C\u002Fa> page.\u003C\u002Fp>\n\u003Ch3>What's included\u003C\u002Fh3>\n\u003Cul>\n\u003Cli>\u003Cstrong>Native macOS app for Apple Silicon\u003C\u002Fstrong> built with SwiftUI and AppKit.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Fully local fine-tuning workflow\u003C\u002Fstrong> for MLX-compatible language models.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>10 training algorithms:\u003C\u002Fstrong> SFT, DPO, FTPO, CPO, ORPO, GRPO, Online DPO, XPO, RLHF Reinforce,\nand PPO.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Dynamic SFT objectives:\u003C\u002Fstrong> NLL, memory-bounded Chunked NLL, and Dynamic Fine-Tuning.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Multiple training modes:\u003C\u002Fstrong> LoRA, DoRA, QLoRA at 4\u002F6\u002F8-bit, full fine-tuning, and\nQuantization-Aware Training (QAT).\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Live training observability\u003C\u002Fstrong> with loss, learning rate, gradient norm, throughput,\nprogress, logs, and recent-step charts.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Memory-aware run planning\u003C\u002Fstrong> with live wired\u002Factive memory monitoring and Resourc\u003C\u002Fli>\n\u003C\u002Ful>\n",1784564568420]