Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

Impediments to changing the dlrmv4 runs count

Open
  • Is something wrong?
#480 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
25/100
Issue type
Refactor
Clarity
Mostly clear
Activity status
Active
Tech stack
python
Domain
tooling

Research direction

Start by comparing the run-count and scoring rules in training_policies/training_rules.adoc with benchmark_meta.py and rcp_checker/rcp_checker.py. Then trace discard-count handling in rcp_checker.py and result_summarys.py, including the version and unet cases named in the issue. Done means the policy and code agree on the intended counts while preserving old-version results; the issue leaves some policy choices unresolved.

Written by the indexing model from the issue text.

Description

in training_policies/training_rules.adoc:

  1. run-count table (~line 565): change the minimum runs from 10 to 12
  2. scoring rule (line 570) says drop one fastest and one slowest. We should generalize to say "drop ceil(N/10) fastest and ceil(N/10) slowest". That will cover the old unet3d benchmark (which had N=40 and drop 4 highest and lowest), as well as all the N=5 and N=10 (where ceil(N/10) works out to drop 1 highest and lowest), and would cover dropping two highest and two lowest when N=12.
  3. The RCP rule requires references to provide 2N convergence numbers, so we either need to give an exception or we need to increase the number of RCP convergence numbers from 20 to 24.

logging

  1. The code does not follow DRY, so we need to modify both benchmark_meta.py _ALL_RESULT_FILE_COUNT and rcp_checker/rcp_checker.py submission_runs (or better: have the rcp_checker import benchmark_meta.py). The contents seem to be identical, so this should be safe.
  2. _ALL_RESULT_FILE_COUNT and submission_runs are not versioned. So once a benchmark has a particular count, we can't change it in future versions without messing up how the checkers and result summarizers work on old versions.
  3. the discard count is hardcoded to 1 all over the place with special exceptions for unet in some places (and bugs in other places where the unet exceptions are just ignored). Fixing this in such a way that old unet results don't change will require hardcoding a special exception for rounds prior to 6.1 at rcp_checker.py:406 (which doesn't handle unet correctly), but the problem at result_summarys.py:564 was introduced in April of 2026 and should be fixed (generalized) because it used to handle unet but then was hardcoded to 1 in April, so broke old unet results.

And additionally: there was a proposal to make the discard asymetric (discard two highest and only one lowest). And adding that to the code requires additional generalization in about a dozen places where it's currently hardcoded that the upper and lower discard count is the same.

Dominant language
Python
Stars
43
Forks
60
Avg merge
2d 14h
Merged PRs (30d)
1

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from mlcommons/logging

All issues in mlcommons/logging

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.