LOCAL_LLM_BENCHMARKS

Local LLM benchmarks

scripts/benchmark_local_llm.py measures one fictional duplicate-adjudication request against an Ollama model that is already installed and already running. It never downloads a model, launches a server, changes Ollama settings, or stores prompts/responses in its output.

Plan a run without network activity:

python scripts/benchmark_local_llm.py --model existing-model

Run one request and save aggregate metrics:

python scripts/benchmark_local_llm.py --model existing-model --execute --output benchmark.json

The output contains endpoint, model, selected profile label, wall-clock time, response byte count, and process peak RSS. A missing server or model exits with code 2 and a skip: explanation; this is an environmental skip, not a reason to download or configure anything automatically.