LoRA / PEFT adapters

What Is A LoRA Adapter?

A LoRA adapter is not a second model. It is a small set of extra weights, trained separately from the base model, that nudge the base model's behavior in a particular direction when combined with it at inference time.

Full fine-tuning produces a new copy of every weight in the model — for a multi-billion parameter model, that is another multi-gigabyte file per variant. LoRA instead trains a pair of much smaller matrices per targeted layer (a low-rank decomposition, hence the name) that approximate the change fine-tuning would have made, without touching the base weights directly. The practical result is what matters here, not the linear algebra: an adapter file is typically single-digit to a few tens of megabytes, it is swappable independently of the base model, and it does not duplicate the base model's weights.

Why This Matters

Cheap specialization: get a specialized assistant, tone, or skill on top of one base model without paying the storage and load-time cost of a full model per variant.

Fast iteration: adapters are small enough to train, share, and swap far more quickly than retraining or redistributing a full checkpoint.

One resident base model, many behaviors: the eventual goal (Phase 2 below) is to keep one base model loaded and switch which adapter is applied per request, rather than loading a different full model for each behavior.

Standard, portable format: adapters trained anywhere in the HuggingFace PEFT ecosystem should work with Reckoning's support, without a Reckoning-specific training or export step.

One prompt, a different behavior

The integration test loads a base model, generates once, activates the adapter, and generates the same prompt again:

ModelFetcher fetcher = new ModelFetcher(
    "unsloth", "Llama-3.2-1B-Instruct");

try (AbstractModel model = AutoModelForCausalLm
        .newBuilder(fetcher).buildLocalTransformerModel()) {
    String baseline = generate(model);
    System.out.println("[base model] " + baseline);

    model.registerLoraAdapter("bunnycore-chatml",
        new LoraAdapterModelFetcher(
            "bunnycore",
            "Llama-3.2-1b-chatml-lora_model"));
    model.setActiveAdapter("bunnycore-chatml");

    String adapted = generate(model);
    System.out.println("[adapter active] " + adapted);
}

The test asserts that the adapted response differs from the base response. It then clears the adapter and verifies that the original base behavior returns.