paper(§7B): inject measured M13 7B rung + honest AXIS-2-only-measured framing#1956
Merged
Conversation
…rs left Table tab:m13-prereg filled with measured values: ce@3499 1.42529 (run low); AXIS-2 GREEN (CE_real 1.90741 < uniform 5.54518 & < shuffle 5.31616, byte-exact clm_decode_mirror on the 7B .clm); AXIS-1/3 GREEN (substrate-native, .clm-indep, scale-invariant = 3B); .clm v0.3 mount GREEN (d6208 config-agnostic admit); val_ce/rel_gap eval-deferred (post-train eval crashed, ckpt saved BEFORE eval by the final-save patch — all 7 ckpts on HF). Flipped PAPER.md milestone + the 'currently training' refs (×5) to 'STEP 3500 complete, measured'. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…S-1/3 substrate-native Committed verdict .verdicts/core-3axis-mount/probe.txt proves AXIS-1/3 are .clm-weights-independent (ran with a d768 .clm, same motiv 0.6700>0 + composed 101>72). Abstract line 137 'measured GREEN' could read as all-three-on-7B-weights; tighten to match the already-honest tab:m13-prereg caption. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…bstract (48pp, lualatex) Co-Authored-By: Claude Opus 4.8 (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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Fills the paper's §7B placeholder with the measured M13 7B rung (STEP 3500 complete), and makes the abstract honest about which axes were measured on the 7B trained weights.
§7B measurement (tab:m13-prereg)
1.90741< ln2565.54518& < shuffle5.31616, byte-exactclm_decode_mirror.pybrain_emit(pure_field),.clm-weights-independent.clm-weights-independent.clmv0.3 mountd6208/L30/E30, config-agnostic decodeWhy AXIS-1/3 are honestly "substrate-inherited", not "7B-measured"
The committed verdict
.verdicts/core-3axis-mount/probe.txtran the probe with a d768.clmand still producedmotiv 0.6700 > 0(AXIS-1) andcomposed 101 > parts 72(AXIS-3). Those axes callbrain_emit(pure_field)and never read.clmweights (CORE/three_axis_probe.hexaL76–110), so they reproduce identically at d768 / 3B / 7B by construction. The abstract previously read "3-axis @ 7B measured GREEN" which could imply all three were measured on 7B weights; tightened to credit AXIS-2 as the measured axis.Honest scope
M13 is a measured scale-extension of the terminal 3B finding, deeply undertrained (
0.0027tok/param,a_scale_honest_scope) — NOT a closure.val_ce/rel_gapare eval-deferred (post-train eval crashed; final-save-robust patch wrote the ckpt before eval, weights safe on HF). All 7 ckpts (500–3500) preserved.Build
lualatex × 3 (Noto Serif CJK KR), 48pp, rebuilt on aiden. PDF byte-identical local↔aiden.
🤖 Generated with Claude Code