Skip to content
Backpropagate

Handbook

Welcome to the Backpropagate handbook. This is the complete operator guide — every flag, every env var, every error code, every recipe — from first install through training, export, and running your fine-tune locally with Ollama. Start at Getting Started if you’ve just installed; jump to Troubleshooting if something is on fire; head to Recipes for paste-and-run snippets matched to common operator workflows.

What’s inside

Section titled “What’s inside”
  • Getting Started — Install and train your first model in under 5 minutes
  • Training — Basic training, multi-run SLAO, callbacks, dataset formats, and model presets
  • Preference tuning (ORPO / SimPO / KTO) — v1.6 — when to use each method, paired vs unpaired data, the VRAM envelope, hyperparameters, and the cited papers
  • Full fine-tuning (mode="full") — v1.4 — when to use full FT, the 4B parameter ceiling, and the LoRA-vs-full quality math
  • Export — GGUF export, Ollama registration, quantization options, and custom Modelfiles
  • Recipes — Common operator workflows — paste-and-run snippets keyed by what you actually want to do
  • Reference — Architecture, CLI overview, GPU safety, checkpoints, Windows support, error handling, and bug-reporting
  • Security — Threat model, GHSA-f65r-h4g3-3h9h auth middleware closure, deployment patterns, and vulnerability reporting
  • Error codes — Catalog of every structured BackpropagateError.code with meaning and recommended fix
  • VRAM estimator — v1.4 — Trainer.estimate_vram() + backprop estimate-vram for pre-flight VRAM sizing
  • Troubleshooting — Symptoms-first reverse index for first-run failures
  • Troubleshooting (CUDA) — Deep-dive on OOM, CUBLAS, NCCL, mixed-precision pitfalls, driver mismatch
  • Environment variables — Every BACKPROPAGATE_* knob, grouped by family
  • CLI reference — Every backprop subcommand and flag
  • Migrations — Operator-facing upgrade paths between Backpropagate versions (v1.5 → v1.6, v1.3 → v1.4, and earlier hops)
  • Beginners Guide — New to LLM fine-tuning? Start here for a step-by-step walkthrough

Back to landing page