docs(results): where the head-row family lost BM, and scorer nondeterminism - #68
Open
opariffazman wants to merge 1 commit into
Open
docs(results): where the head-row family lost BM, and scorer nondeterminism#68opariffazman wants to merge 1 commit into
opariffazman wants to merge 1 commit into
Conversation
…minism Diff of the four DPO arms' scored answers, 900 task-instances, no GPU. The head-row tools are net -15 of the -39; find, the diluted one, is +10. 60% of losses keep the same tool and 64% are longer: elaboration, not tool confusion. Separately, 13 of 900 groups score a byte-identical command both 0 and 1, which accounts for most of issue #57's noise floor. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Answers the "next lever" line at the end of the #51 section. Analysis only — no model, no GPU, no dataset change.
The hypothesis in RESULTS was inverted. Net flow v7-dpo → v11-dpo by gold-side tool, three registers pooled:
The head rows doubled their share of the pool (5.1% → 9.8%) and the model got worse at exactly those tools.
find, diluted 5.03% → 4.86%, is net +10. So dilution is not the mechanism.60% of losses keep v7-dpo's tool, 64% are longer, mean command 6.79 → 7.55 tokens. They read as elaborations:
chown→sudo chown,df -h | awk→… | cut -d% -f1,free -h→smem --system.Separately: the ALFA scorer is nondeterministic. 13 of 900 task-arm groups score a byte-identical command both 0 and 1 — 1.4%, which is issue #57's noise floor (0.013/register) almost exactly. #57 read that as training noise; most of it is the embed scorer at threshold 0.75. Worth knowing before any future CI is called tight.
Suggested next steps are in the section; neither is worth a run on its own.
🤖 Generated with Claude Code