llamero: Skill for Claude Code

Training adapters with llamero (LoRA / QLoRA)

llamero trains LoRA adapters in-process on the resident model: build a golden dataset of prompt/completion pairs, call train_adapter, watch the loss stream down, then activate the adapter. The knowledge behaves like a removable filter — on when activated, gone when deactivated, and the base model never reloads at any point.

When the base model is quantized (any *-4bit model — the normal case), the adapters automatically train QLoRA-style over the quantized layers. There is no separate QLoRA mode to configure.

Prerequisites are the same as the local-inference skill: Apple Silicon, the Swift bridge built once via build.sh, and a loaded model session.

Storage root

Default adapter storage is under ~/.llamero. Apps that need app-owned AI data set this at boot before creating runtimes or training:

Llamero.storage_root = Path.home.join(".scribe")

LLAMERO_HOME=/path/to/root is the env alternative; programmatic wins. The default train_adapter output becomes Llamero::Storage.adapters_dir/<name>/.

Recipe: teach a model facts it cannot know (complete program)

require "llamero"

# Train on a DENSE model (see "Choosing a base model" below).
runtime = Llamero::Native::MLXRuntime.new(model_id: "mlx-community/gemma-3-1b-it-4bit")
session = runtime.start_session
session.load_model

# 1. Golden dataset: prompt/completion pairs. Use several phrasings of each
#    fact so short training runs generalize beyond exact wording.
#    No format: needed — train_adapter automatically uses the model's own
#    chat template (or the right built-in) for datasets with the default.
dataset = Llamero::Native::TrainingDataset.new(
  system_prompt: "You are a Crawley LX-900 bulldozer maintenance expert."
)
dataset.add("What fuel injectors does the Crawley LX-900 use?",
  "The Crawley LX-900 uses BR-7741 fuel injectors rated at 2,150 PSI.")
dataset.add("Tell me the LX-900 fuel injector part number.",
  "The part number is BR-7741, rated at 2,150 PSI.")
dataset.add("What oil does the Crawley LX-900 use?",
  "The LX-900 uses 15W-40 heavy duty diesel engine oil.")

# 2. Hyperparameters. lr 1e-4 memorizes small fact sets quickly.
config = Llamero::Native::AdapterTrainingConfig.new
config.iterations = 300
config.learning_rate = 1e-4

# 3. Train. Streams live loss. Speed scales with model size: ~1 minute for
#    300 iterations on a 0.6B model, tens of minutes on a 2B-class model.
descriptor = session.train_adapter("lx900-manual", dataset, config) do |progress|
  puts "iter #{progress.iteration}/#{progress.total_iterations}: loss=#{progress.loss.round(3)}"
end
puts "saved to #{descriptor.path}"

# 4. The adapter is auto-registered. Toggle the knowledge on and off:
session.activate_adapters(
  Llamero::Native::AdapterStack.additive([
    Llamero::Native::AdapterSlot.new("lx900-manual"),
  ])
)
puts session.chat([Llamero::Message.user("What injectors does the LX-900 use?")]).content
session.deactivate_adapters

runtime.close