Fine-tuning, Inference, and Evaluation for LLM tool calling with NVIDIA NeMo Microservices

Why it matters

When prompt changes cannot fix wrong tool selection, a training candidate is the next option. This tutorial runs that candidate end to end and evaluates it before use. Its NeMo Microservices platform docs have moved to NVIDIA’s archive, so treat the deployment steps as legacy.

What to practise

Work through the four notebooks: prepare the Salesforce xLAM function-calling data, run a LoRA fine-tune of llama-3.2-1b-instruct with NeMo Customizer, evaluate tool-calling accuracy with NeMo Evaluator, and add guardrails. Compare the base and customized models on function_name_accuracy and function_name_and_args_accuracy. Note that the workflow works only with a NIM for inference, and that a tool schema missing items for arrays or properties for objects, or a function with more than eight parameters, stalls tool calls.

How it fits

This is 8.4.2’s supervised fine-tuning path, applied to tool calling. 8.4.1 decides whether a failure warrants training at all, and 8.4.3 covers promoting the resulting LoRA adapter with a rollback.