We publish each stage of NewsShorts-Bench as a long write-up: the question, the systems we ran, a results table, the method and what broke. Twelve stages in, the tables were only readable as web pages. Nobody could load them into pandas, and nobody could cite them properly.
So we archived the results as a dataset with a DOI. It took one afternoon. Here is the whole process, because most of it applies to any project that publishes numbers in Markdown.
1. Export from the pages you already publish
Our write-ups are MDX files with the results in frontmatter (name, value, confidence interval, extra columns). The rule we kept: the dataset must be generated from the same files the site renders, never retyped. If a number changes, both change.
import csv, glob, yaml for path in sorted(glob.glob("content/benchmarks/*.mdx")): fm = yaml.safe_load(open(path, encoding="utf-8").read().split("---")[1]) keys = [c["key"] for c in fm.get("columns") or []] with open(f"results/{fm['slug']}.csv", "w", newline="", encoding="utf-8") as fh: w = csv.writer(fh) w.writerow(["system", "kind", fm["metric"], "ci_low", "ci_high", *keys]) for r in fm["results"]: stats = r.get("stats") or {} w.writerow([r["name"], r["kind"], r["value"], r.get("ciLow"), r.get("ciHigh"), *[stats.get(k) for k in keys]]) Plus one index.csv with a row per stage: question, verdict, metric, method, hardware, run date and the URL of the full write-up. The write-up stays the source of truth for method and caveats; the dataset points back to it.
2. A README that a stranger can use
Three things mattered:
- what each file is, in one line each
- where the numbers came from, and what is not included (we do not redistribute the source videos)
- the license (CC BY 4.0) and how to cite
3. Zenodo for the DOI
Zenodo is free, run by CERN, and mints a DOI on publish. Our notes from doing it:
- You can log in with GitHub. Use an organization or project account, not a personal one, if the dataset belongs to a project.
- Add the author as an Organization, not a person, when the work is a collective's.
- The description field is a rich-text editor. Link the write-ups from it, so readers land on the method.
- Files are frozen once published; metadata can still be edited. Check the zip before you press Publish twice.
4. Close the loop on your own site
A DOI nobody can find does little. We added a "Download and cite" block to the benchmark index with the citation, and schema.org/Dataset markup (identifier = the DOI URL, license, creator, distribution). Search engines and AI answer engines read that markup, and it ties the page and the archive together.
What it cost
The exporter is about 30 lines. Most of the time went into the README and checking that every CSV matched its page.
The dataset: NewsShorts-Bench results on Zenodo (doi:10.5281/zenodo.24324094). The write-ups behind it are on Every Local AI.
If you publish benchmark numbers in blog posts, try archiving them this way. Tables in Markdown are easy to read and hard to reuse; a CSV with a DOI is the opposite, and you can have both.


