Getting support¶
Where to ask¶
| Channel | Use for |
|---|---|
| MLPerf Endpoints working group | Joining the group — the primary route for rule questions and proposals |
endpoints@mlcommons.org |
General questions about the benchmark and submitting |
mlcommons/endpoints_policies issues |
Rule ambiguities, errors in the rules, v1.0 proposals |
mlcommons/endpoints-submission-cli issues |
Bugs in the submission CLI or checker |
mlcommons/endpoints issues |
Bugs in the reference benchmark client |
| Your submission's review thread | Anything about a submission already in review |
No response times are published
None of the sources state a support SLA, or a turnaround for PRISM API-creation access. Build slack into your plan rather than assuming same-day answers.
What to include when asking¶
A question with these attached gets answered once instead of three times:
- Which division and scenario you are submitting under
- Versions — the
endpointscommit SHA you built from, andendpoints-submission-cli --version - The exact error text, not a paraphrase
- The checker output —
submission-checker check … --output checker.json - The relevant
point.yamlwith secrets removed - What you expected, and what the rules clause you are reading says
For a run problem, report.txt and the non-histogram fields of result_summary.json are usually
enough. Don't attach events.jsonl, which runs to hundreds of megabytes.
Never paste your PRISM token
Not in an issue, not in a log, not in a config attachment. If you think one has leaked, rotate it from the API Keys dashboard immediately. It can be used to withdraw your submissions.
Questions this documentation cannot answer¶
Some things are genuinely not published anywhere. If your question is one of these, go straight to MLCommons rather than searching further:
- Whether your organisation needs MLCommons membership to submit
- The PRISM and Member Central URLs
- Which seed set a v1.0 submission should bind
- The v1.0 supported model list, accuracy targets, and dataset paths
- CoN client locations and scheduling
- The Preview Availability Tracker and public results URLs
Full list with context: Open questions.
During review¶
Objections are filed and resolved on the submission's review thread, not through support channels. Meetings are convened only when an objection escalates, a party explicitly requests one with a written agenda and specific questions, or the review chair decides a meeting would materially accelerate resolution.
If you think a reviewer has a conflict of interest, raise it with the review chair. You can do this at any time during review, and the chair decides whether they should step aside.
If you believe a published result involves fraud or misrepresentation, that is a separate route: any MLCommons member may raise it by email to any MLCommons working-group chair. It carries no time limit. Concerns that do not allege fraud are handled as late objections or audit nominations instead.
Reporting a documentation problem¶
This site is a guided path over the authoritative sources, and it can be wrong in ways the sources are not.
- If this site contradicts the policy repository, the policy repository is correct and this site has a bug. Please report it.
- If a page is out of date, check its
Last verified against:line. Pages are verified against a specific commit, and the upstream rules are a live development branch. - If something was hard to find, that is also a bug worth reporting. A question that keeps being asked means a page failed to answer it.