Throughput and output-correctness benchmark data for local LLM inference on a repurposed enterprise workstation

Complete measurement data behind the paper "Reusing Obsolete Windows 10 PCs for On-Premises Large Language Model Inference" (I. Curington and K. Lano, 2026, Frontiers in Computer Science). The study asks whether a decommissioned enterprise workstation, retired because it cannot run Windows 11, is capable of useful local large language model inference. Answering that needs two kinds of measurement usually reported separately: how fast a model runs, and whether its output is correct. A fast wrong answer has no value, so both axes are published together here.

The deposit holds one inference-throughput run and the per-item outputs of three correctness benchmarks (GSM8K, HumanEval, TruthfulQA) across nine model configurations: gemma4:e4b, gemma3:12b, qwen3:14b, granite4.1:8b, phi4:14b, deepseek-r1:1.5b, mistral-nemo:12b, gpt-oss:20b, and mistral-nemo:12b repeated at Q4_K_M as a quantisation control. Every number in the paper's tables and figures can be recomputed from these files using the published scripts.

Data is released per item rather than as summary statistics, because the paper makes claims about failure modes — models that fill the context window with reasoning before answering, answers truncated by a stop sequence, code that fails extraction rather than execution — that can only be checked against raw outputs. Files comprise the throughput CSV, results JSON and samples JSONL per model and benchmark, five adjudication files supporting the human validation of the LLM judge, and MANIFEST.json recording per-file provenance.

Measurements come from two machines. Throughput was measured on a Dell Precision T7810 with an NVIDIA RTX 3060 (12 GB VRAM), since retired; correctness was measured on a separate Windows 11 workstation fitted with the same GPU model. Correctness depends on model weights and sampling parameters rather than on the GPU, so the two are comparable, but they were not measured on one machine.

Limitations a reuser should know: throughput is a single run of five repetitions per prompt; correctness scores come from one greedy (temperature-0) generation per item; all runs used Ollama's default 4,096-token context window, which clamps generation regardless of the max_tokens values recorded in the scripts; TruthfulQA answers in the four runs scored under the stock truthfulqa_gen task are truncated at the first paragraph break; and quantisation levels are each model tag's default, so they are not uniform across models.

Code: https://github.com/t3dr0/ollama_benchmark — the benchmarking suite, the custom chat-aware HumanEval task, and the scripts that turn this data into the paper's tables and figures. Archived at DOI 10.5281/zenodo.22079049. The data here is CC BY 4.0; the code is AGPLv3 and that licence does not apply to this data.

v2 (2026-09-14). Corrects the deposit README. Generation budgets requested via max_tokens are clamped by Ollama to the context window, which was left at its 4,096-token default for all runs; the README previously described these as exhausted generation budgets. The TruthfulQA truncation caveat is now scoped to the four stock-task runs rather than implied of all nine. No data file has changed — all 62 files are byte-identical to v1.