feat(bench): add --compare-static fixed-depth baselines to mtp-adaptive - #276
Open
rinaldofesta wants to merge 1 commit into
Open
feat(bench): add --compare-static fixed-depth baselines to mtp-adaptive#276rinaldofesta wants to merge 1 commit into
rinaldofesta wants to merge 1 commit into
Conversation
Runs the same prompt suite at listed fixed depths after the adaptive pass (AdaptiveDepthPolicy with min == start == max), emitting static_rows and static_summaries next to ar_rows, with exact_match_vs_adaptive at greedy. Output unchanged unless the flag is passed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owner
|
Useful knob, fixed depth baselines belong in the harness. Reviewing. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What is the current behavior?
mtp-adaptivecan compare an adaptive run against AR (--compare-ar), but not against fixed depths on the same suite. Answering "does the adaptive policy beat the best static depth?" requires separate invocations with--min-depth/--start-depth/--max-depthpinned to the same value and a hand-merge of the outputs. That workaround is what produced the numbers in #271.What is the new behavior?
--compare-static 2,3runs the same prompt suite at each listed fixed depth after the adaptive pass, reusing the existing generation path (AdaptiveDepthPolicywithmin == start == max, fresh policy per case, same seeds, sampler and verify settings).static_rows(per-depth per-prompt rows: tok_s, acceptance_by_depth, drafted/accepted) andstatic_summaries(per-depth mean_tok_s and draft totals) next to the existingar_rows.temperature <= 0each static row recordsexact_match_vs_adaptive, mirroring the existingexact_matchsemantics vs AR.(): output is byte-identical unless the flag is passed.Validation
python -m py_compileon both files.--policy expected_value --max-depth 3 --limit 2 --compare-static 2,3:compare_static: [2, 3]recorded,static_summariespopulated (D2 21.98 tok/s, D3 22.22, adaptive 21.02 on the 2-prompt subset),exact_match_vs_adaptive: nullat temp 0.6 as intended.Motivated by the discussion in #271: a same-suite static baseline makes the adaptive bench self-contained.
Implemented with the help of a coding agent (Claude Code) and reviewed locally by the author.