====== LET'S DEFINE THE PIECES ====== Everything you need to know falls into one of three categories: ---- ===== 1. MODELS (LLMs) ===== ---- * These are the brains. \\ * They are NOT software. \\ * They are NOT programs. \\ * They are NOT plugins. \\ * They are neural networks stored in a single file, usually: \\ **model.gguf \\ model.safetensors \\ model.bin** \\ \\ \\ **Examples of models:** \\ LLaMA-3 8B \\ LLaMA-3 70B \\ Mistral 7B \\ Mixtral 8x7B \\ Qwen 2 7B \\ Phi-3 \\ \\ **A model file contains:** \\ Neurons \\ Synapses \\ Every learned pattern \\ All the intelligence \\ \\ **A model does NOT:** \\ * Have a UI \\ * Open PDFs \\ * Connect to NASA \\ * Run indexing \\ * Provide a chat window \\ \\ It only takes text in → text out. \\ ---- ===== 2. MODEL RUNTIMES (ENGINES) ===== ---- These are the programs that LOAD and RUN the model's brain. \\ Think of a “runtime” as the machine that runs a model file. \\ \\ **===== Runtimes include ===== :** \\ **✔ Ollama** * Terminal-based * Local API * Can fine-tune * Good for automation * Good for pipelines * Very flexible * Acts like a backend server **✔ LM Studio** * GUI desktop app * Easy model downloading * Drag-and-drop PDFs * File chat * Rudimentary RAG * Easy to test many models * Great for tinkering **✔ GPT4All** * GUI * Also a runtime * Similar to LM Studio * Not as modern **✔ koboldcpp** * Runtime specialized for story-writing/roleplay * GUI * Some fine-tuning tools ✔ Faraday / Unsloth / Axolotl (training tools, not chatbots) These are **training engines**, not chat apps. You don’t talk to them; you use them to train a model. ==== Summary ==== **Models** = The brains \\ **Runtimes** = The engines that run the brains \\ **Training tools** = The machines that modify the brains \\ ---- ===== 3. TRAINING TOOLS (NEURAL LEARNING, “NEW SYNAPSES”) ===== ---- This is where real learning happens. Not prompts. Not memory. \\ Not RAG. (library model), Actual neural adjustment. \\ \code\ Tools include: **✔ Axolotl** * THE standard for fine-tuning LLaMA, Mistral, etc. * Used by researchers, labs, and giants. **✔ Unsloth** * A faster, more GPU-efficient training toolkit. * Great for consumer GPUs like yours. **✔ Faraday** * A more experimental training tool, but advanced. \\ ** * These tools: ** * Read your dataset * Modify the model’s weights * Produce a new personalized model This is real learning.