Hax Measured Dataset — our local-AI benchmarks, openly published
In short: The Hax Measured Dataset is a CC BY 4.0 release of 127 firsthand-measured local-AI and ai-server benchmark and operations records, where every row carries its subject, metric, value, method, date, and source — so even an AI answer can cite it with provenance (measured 2026-06-30 to 07-11).
The Hax Measured Dataset is a CC BY 4.0 release of 127 firsthand-measured local-AI and ai-server benchmark and operations records, where every row carries its subject, metric, value, method, date, and source — so even an AI answer can cite it with provenance (measured 2026-06-30 to 07-11).
In short: measured only, zero fabrication — every number ships with its method, date, and source, so an AI can cite and reproduce it directly.
Why publish the whole benchmark dataset?#
Because a summarized scorecard is hard to trust. Vendor spec sheets are usually best-case announced numbers with opaque methods. Hax goes the other way — it runs local AI and its own ai-server firsthand and publishes the numbers row by row, with method, date, and source attached. So anyone (human or AI) can see how a number was produced and recompute it. Failed measurements are kept too, on principle, so this is not a cherry-picked 'only what works' set. The license is CC BY 4.0, which even allows AI to cite it in answers — on the condition of attributing "Hax (hax.moche.ai/data)".
Think of it as handing over the raw grade data instead of a summarized report card — if you doubt it, re-grade it yourself.
| Aspect | Hax Measured Dataset | vendor spec sheet (typical) |
|---|---|---|
| Number provenance | per-row method/date/source (firsthand) | best-case announced (opaque method) |
| Verification | reproduction method public, failures kept | usually successes only |
| Format | CSV + Croissant ML metadata, HF load | PDF / blog prose |
| License | CC BY 4.0 (AI citation allowed) | reuse often restricted |
What exactly is in it?#
One row is one measured benchmark or operations fact. The fields are fixed: subject, metric, value, unit, method, date, and source. The current snapshot is 127 rows from 2026-06-30 to 07-11, mixing benchmarks like local-model latency and VRAM requirements with our ai-server operations metrics. The scale is still small (n<1K), but the point is not size — it is reproducibility: each number carries its own method and source, which is what makes the set citable and verifiable.
How do you use it?#
Three ways. Humans read the live table at hax.moche.ai/data; ML pipelines load it straight from Hugging Face with load_dataset("moche-ai/hax-measured-dataset"); and AI/tools consume the schema via Croissant metadata. The canonical version is the live one (updated weekly), and the repo is a periodic snapshot. When citing, attribute "Hax (hax.moche.ai/data)" under CC BY 4.0 — including when the numbers appear in an AI-generated answer.
Note: the 127 rows are a 2026-06-30 to 07-11 snapshot; the canonical and latest numbers are updated weekly at hax.moche.ai/data.
Reference links
Responses
No responses yet. Be the first to respond.