mlabonne/llm-course

This page introduces an external open-source repository. It is not an HDATF product.

What it does

This is a course for people getting into large language models (LLMs), kept in one GitHub README with roadmap images, concept notes and resource links. Its three parts are optional fundamentals (math, Python, neural networks), The LLM Scientist on model building, and The LLM Engineer on applications and deployment. The Scientist track covers architecture, pre-training, post-training data, fine-tuning, preference alignment, evaluation, quantization and model merging. The Engineer track covers running models, vector storage, retrieval-augmented generation (RAG), agents, inference optimization, deployment and security. Hands-on notebooks for fine-tuning, quantization and merging are linked on Google Colab outside the repository; some descriptions name RunPod, the Hugging Face Hub or ZeroGPU. The repository holds only the README, a license and images, so nothing needs installing; the README says the course will stay free. Notebook contents, GPU requirements and linked materials' terms were not inspected. Notebook titles target models such as Llama 2, Llama 3.1 and Mistral-7b, so library versions should be checked first.

License

Apache-2.0 Permissive, with a patent grant. Commercial use is allowed; keep the notices and state your changes.

More in this category

• Gathin23/tv-vibehack
  This prompt plus scripts kit, with a native launcher, lets TV owners disable ad services, telemetry and preloaded apps on Android TV and Google TV. The README says Claude Code did the original work on a real TV with PROMPT.md. Once USB debugging is on and the computer is approved on the TV, a coding assistant with shell access follows the prompt and asks which apps to keep. Over ADB (Android Debug Bridge) it backs up packages and settings, disables apps with pm disable-user and checks each step with TV screenshots. The tools/restore.sh script re-enables those packages and the stock launcher. The Java launcher lists installed apps, pins four to a dock, switches HDMI and Live TV inputs and lacks internet permission. It builds with JDK 17 and Android SDK 34; manual tools need Python 3 with adb-shell. Five tested model lists are included; Samsung, LG and Fire TV are unsupported. Most lists also disable the update service, so firmware updates stop until it is re-enabled.
• lucannez64/darling-lol
  This is an experimental collection of Darling runtime patches and integration tools for developers testing the macOS League of Legends client on x86-64 Linux. It addresses graphics, audio, input and startup compatibility rather than providing a finished game launcher. Inputs include pinned Darling sources and a separately obtained official game installation; the intended result is a patched runtime that can launch the client and game. The repository contains source patches and build, installation and regression-check tools, but does not include game files, account databases or tokens. Requirements include Linux 5.14 or newer, X11 or Xwayland, hardware Vulkan/OpenGL drivers, build dependencies and Python 3.12 or newer. Building the runtime and installing the game are separate tasks, so downloading this repository alone does not make the game usable. The author reports limited maintenance and says a fresh checkout has not been validated end to end on another machine. The README explicitly warns of permanent account bans; compatibility work is not Riot approval, and the reported performance was not independently verified here.
• darlinghq/darling
  Darling is a compatibility runtime for developers seeking to run macOS programs on Linux without a virtual machine. It takes macOS executables and their required files and provides a Linux-hosted environment intended to run them. The official README describes a Mach-O loader, a userspace darlingserver and reimplemented Darwin interfaces and Apple frameworks; the underlying source was not independently audited here. Some command-line tools work, while graphical application support remains under development and Xcode itself is not yet supported. Users can begin with an available official package or build the runtime with its documented dependencies, then prepare the application separately. The build guide requires x86-64 Linux and tools such as Clang and CMake, together with submodules and supporting libraries. DPREFIX keeps a macOS-like filesystem, but its overlayfs use restricts locations such as NFS or eCryptfs. Installing the runtime is not a promise that a particular application will work, and community packages are explicitly not vetted by the project.
• chenglou/handwired
  Handwired is an MIT-licensed TypeScript project that writes program logic in the form of a neural net. A program is a fixed list of operations over 32-bit floats, mostly weighted sums and ReLU products plus 0-or-1 rounding and table reads, and its weights are written by hand without training. The README presents it as turning regular JavaScript into a neural net. The docs state that no tool converts existing JavaScript source, and programs are written in builder code that appends operations to a Circuit. An if-else becomes arithmetic in which both branches are computed and a value of 0 or 1 picks one, and the README says this keeps the whole program differentiable, while the docs require explicit gradient options for rounding and table reads. The program reads one Float32Array of cells, and host JavaScript writes inputs into it, runs an evaluation, copies the next state back and draws the results in the DOM. Programs run in JavaScript on the CPU, and only the JSON search demo turns them into WebGPU shaders to search 100,000 generated JSON documents at each keystroke. The demos are HandwiredOS, with three windows and a dock run by one program, a photo gallery whose layout network is trained with the program's gradients, and the JSON search page. Getting started needs Bun, with 1.4.3 as the tested version, where bun install and bun start open the demo site and a toggle switch template is the start of a new app. The lab status page lists known issues such as gradients that change with the order of code statements, and the lab folder holds older experiments.
• nyacg/do-on-container
  do-pool is a TypeScript experiment in the repository do-on-container. It runs the same Cloudflare Durable Object classes either as real Durable Objects or hundreds per container inside workerd on Cloudflare Containers. It targets objects that stay awake but do little, such as game loops, presence, upstream feeds and AI agents waiting on models, which are billed for every second they stay awake. In POST.md, 1,000 such objects cost about 22 dollars a month in the pool and about 4,050 dollars as real Durable Objects, and objects that mostly hibernate are said to fit real Durable Objects better. A placement object assigns objects to containers, scales nodes from CPU and connection statistics, moves several objects at a time off busy nodes and uses per-assignment epochs against stale moves. Every 30 seconds each node writes all changed objects as one R2 segment plus a manifest, and on SIGTERM the node saves everything first. On a hard failure up to 30 seconds of writes can be lost, and in the write-up's SIGKILL test 5 of 10 writes survived while all 10 survived a drain. Container location cannot be chosen, and in the write-up's browser measurement pool round trips took about 80 ms against about 20 ms for real Durable Objects. Running it needs a Cloudflare account and API token, an R2 bucket and Docker to build the linux/amd64 image, and status endpoints are public while only admin calls need a token. The write-up is marked as a draft, the repository has no license file, and the two demos are live cursors and a Yjs collaborative editor.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.