TechFuelHQ RTX 5080 local-LLM throughput dataset: raw capture files Dataset page: https://techfuelhq.com/data/rtx-5080-llm-throughput/ Published CSV: https://techfuelhq.com/data/rtx-5080-llm-throughput.csv Data license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) capture-2026-08-12T061545.json capture-2026-08-12T061806.json source of 12 CSV rows capture-2026-08-12T084340.json source of 4 CSV rows capture-2026-08-26T021714.json capture-2026-08-26T022218.json source of 16 CSV rows capture-2026-08-26T022548.json source of 2 CSV rows llm_throughput_capture.py the capture script (method llm-throughput-ollama-v1) Each of the 34 CSV rows matches one cell in the files marked above on date, model, context, prompt tokens, median decode rate and median prefill rate. The 061545 and 021714 files are earlier runs of the same cells on the same day; the CSV does not use them. All six ran on ollama 0.32.1, driver 610.88. The 2026-08-12 captures ran an earlier revision of the script without the retry on a transient ollama HTTP 5xx. This copy differs from the revision that ran on 2026-08-26 only in how it finds its output folders: outside the site repository it writes under the directory you run it from (raw JSON to ops/audit/evidence//llm-throughput-rtx5080/, the series CSV to ops/data-series/ and static/data/). The measurement code is unchanged. Needs ollama on 127.0.0.1:11434 with the models already pulled, and nvidia-smi on PATH for the GPU, driver and VRAM fields. Standard library only; tested with Python 3.13. python llm_throughput_capture.py --models gpt-oss:20b qwen2.5:14b --contexts 4096 16384 --prompt-tokens 512 2048 --reps 3