[ShriyaRishab]

New Round: v6.1

Round chair: @ShriyaRishab
Prior round: v6.0
Submission deadline: Oct 16, 2026
Ruleset: 6.1.0

Follow-up PRs from this round rollout (nothing merged; please review first):

• training: Introduce MLPerf Training v6.1 round (retire 405B and DLRM-DCNv2) training#901
• logging: already on mlcommons/logging master (training_6.1.0/, qwen35, dlrmv4) — no additional PR
• policies: Add Training v6.1 column to the compatibility table policies#224

Website / marketing updates are out of band (email marketing@mlcommons.org).

Confirm round scope with the Training WG

• Finalize benchmark list for v6.1: carrying forward, new, and retired benchmarks
  • Carry forward: flux.1, llama31_8b, llama2_70b_lora, gpt_oss_20b, deepseekv3_671b
  • New: qwen35_397b_grpo (LLM post training), dlrm_v4_hstu (recommendation / Yambda)
  • Retired: llama31_405b, dlrm_dcnv2
• Confirm submission deadline and ruleset version (6.1.0, deadline Oct 16, 2026)
• Confirm any remaining benchmark-specific rule changes that carry into the new round (dlrm_v4_hstu quality target / eval / run count still TBD)

mlcommons/training — reference implementations

• Move retired benchmark folders under retired_benchmarks/ (#901)
• Add a new round section to README.md (v6.1 section already existed; #901 adds DLRMv4)
• For new benchmarks: reference code already on master (llm_post_training, recommendation)
• For carrying-forward benchmarks: confirm reference code still matches the closed-division rules for the new round

mlcommons/training_policies — rules document

• Update document date at the top of the file (Jul 24, 2026)
• For carrying-forward benchmarks: set Latest version available to v6.1 in all active tables:
  • Benchmarks / datasets
  • Closed division models & targets
  • Hyperparameters
  • Quality / evaluation frequency
  • Minimum number of runs
• For new benchmarks: add rows to each table above plus benchmark-specific rules (qwen35_397b_grpo; dlrm_v4_hstu rows present with TBD where not finalized)
• For retired benchmarks:
  • Remove from active tables
  • Move rules to Appendix: Deprecated benchmarks
  • Add a v6.1 appendix section if the round introduced round-specific exceptions (none identified yet)

mlcommons/logging — checkers & summarizer

Already on master (no new PR from this rollout):

• Add training_6.1.0/ compliance-checker configs
• Add or update RCP JSON files under mlperf_logging/rcp_checker/training_6.1.0/
• Update mlperf_logging/benchmark_meta.py
• Add scripts/verify_for_v6.1_training.sh
• Update mlperf_logging/result_summarizer/config.yaml
• Tag a new logging release once you are ready for submitters to pin it

mlcommons/policies — compatibility table

• Add a column for Training v6.1 (#224)
• Mark supported benchmarks for the new round
• Mark all retired benchmarks as unsupported in the new column

Website & communications

• Email marketing@mlcommons.org to update the Training benchmarks page
• Publish benchmark blog posts for new or materially changed benchmarks
• Present round changes at a Training WG meeting

This PR's code change

• Bump active tables from v6.0 to v6.1
• Add LLM post-training (qwen35_397b_grpo) rules
• Add dlrm_v4_hstu recommendation rows (quality/runs still TBD)
• Deprecate llama31_405b and dlrm_dcnv2 into the appendix

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[RissyRan]

LGTM, thanks!

[ShriyaRishab]

@carualejandro please approve and merge.

[ShriyaRishab]

@pavanky please approve