results: GLM-5.3 (full) on SadeedDiac-25 — 9.8971 zero-skip; 5.x regression is family-wide - #71
Merged
Merged
Conversation
…regression is family-wide Both 5.x variants measured: Flash 8.7978 (wrong haraqat 10.05%), full 5.3 9.8971 (missing 3.75%, wrong 8.59%) — vs GLM-5.2's 2.6911 whose wrong-rate (2.64%) matches our dedicated 580M teacher (2.62%) to within 0.02pp. Flash beats the full model significantly (bootstrap CI excludes zero); both decisively behind 5.2 (CI [-18,028, -14,685] error positions, n=1,200). Convention effect 0.015pp — negligible again. 41 exhausted-retry sentinels were caught by the #65 resume guard and re-fetched; final pass has zero empties.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Completes TODO.training-work/01. glm-5.3 at reasoning_effort=low (same 400/1210 thinking rejection as Flash): 9.8971/7.8219 zero-skip, 9.9760/7.9285 raw, 1,200/1,200 with zero empties (41 retry sentinels caught by the #65 guard and re-fetched). Attribution: the two 5.x variants regress on different axes — Flash wrong-haraqat 10.05 percent, full-5.3 missing 3.75 percent with wrong 8.59 — while GLM-5.2's 2.64 percent matches our dedicated 580M teacher's 2.62 percent to within 0.02pp. Bootstrap: Flash significantly better than the full model; both decisively behind 5.2. Ledger row included.