小さなLLM（Llama-3.2-1B）をQLoRAでファインチューニングしてFunction Callingを覚えさせてみたThis article demonstrates fine-tuning the compact Llama-3.2-1B model with QLoRA…

Function calling, the ability of a language model to recognize when an external tool or API should be used and to emit a structured request describing that call, has become a foundational building block for agents and assistants. A recent Japanese-language blog post walks through teaching this behavior to Llama-3.2-1B, one of the smallest models in Meta's Llama 3.2 release, using a technique called QLoRA. The exercise is notable because tool-use behavior is often associated with large frontier models, yet the author reports that a one-billion-parameter model can acquire a usable version of it on comparatively modest hardware.

The core idea rests on two established methods. LoRA, or Low-Rank Adaptation, freezes the original model weights and instead trains a small set of low-rank matrices that are injected into selected layers. Because only these adapter matrices are updated, the number of trainable parameters drops by orders of magnitude compared with full fine-tuning. QLoRA extends this by quantizing the frozen base model to 4-bit precision, typically using the NF4 data type, so the large base weights consume far less memory during training while the LoRA adapters remain in higher precision. The combination makes it feasible to fine-tune a small model on a single consumer or free-tier cloud GPU, which is likely the central appeal for readers working with limited resources.

Function calling itself is fundamentally a formatting and intent problem. The model must learn to read a set of available tool definitions, usually expressed as JSON schemas describing function names, parameters, and types, and then decide whether a user request warrants a tool call. When it does, the model should produce a well-formed structured output, commonly JSON, that names the function and supplies argument values matching the schema. Teaching this to a base model requires training data that pairs user prompts with the correct tool-call outputs, including examples where no tool is needed. The article's approach appears to follow this pattern: curate or format a dataset of instruction-and-tool-call pairs, apply a consistent prompt template, and fine-tune the adapters so the model reliably emits parseable calls rather than free-form prose.

The practical toolchain for this kind of work has matured considerably. The Hugging Face ecosystem, including the Transformers library, the PEFT library for parameter-efficient methods, the bitsandbytes library for quantization, and the TRL library for supervised fine-tuning, provides most of the necessary components. Projects such as Unsloth and Axolotl further streamline the process and can reduce memory use and training time. For inference and deployment, frameworks like llama.cpp, Ollama, and vLLM allow a fine-tuned small model to run locally, which aligns with the growing interest in on-device and privacy-preserving language models.

Context around the base model helps explain why 1B is an interesting target. Meta's Llama 3.2 family introduced compact text models at roughly 1B and 3B parameters, alongside larger multimodal variants, positioned explicitly for edge and resource-constrained scenarios. Small models of this scale trade raw capability for speed, lower cost, and the ability to run without a data-center GPU. Many of the newer instruction-tuned releases from other vendors already support function calling out of the box, but demonstrating that the capability can be added to a small base model through fine-tuning is a useful illustration of how far parameter-efficient methods have come.

Readers should weigh the demonstration with appropriate caution. A 1B model fine-tuned on a specific dataset is likely to perform best on tool schemas and phrasing that resemble its training examples, and it may struggle with unfamiliar tools, complex multi-step calls, or edge cases where deciding not to call a tool is the correct choice. Evaluation matters here: measuring how often the model produces valid, schema-conformant JSON and selects the right function is more informative than anecdotal successes. Reliability, hallucinated arguments, and generalization remain open questions that any production use would need to address.

Even so, the broader takeaway is meaningful. The post reinforces a trend in which capabilities once reserved for the largest systems become accessible on small, locally runnable models through techniques like QLoRA. For developers exploring agents, offline assistants, or cost-sensitive deployments, it offers a concrete, reproducible starting point rather than a purely theoretical one.