verdict(7B AXIS-2): MEASURED on summer — CE 1.90741 GREEN + memory-bounded lazy mirror#1962
Merged
Merged
Conversation
…unded lazy mirror Genuine 7B decode forward over the serialized .clm v0.3 bytes (d6208/L30/E30, ~7.057B) on summer: CE_realtext 1.90741 < uniform ln256 5.54518 AND < shuffle 5.31616 => AXIS-2 GREEN. Confirms the paper's 1.90741 (was previously unbacked by a committed verdict). Tool: clm_decode_mirror_lazy.py — byte-IDENTICAL math to clm_decode_mirror.py but lazy per-block int4 dequant + vectorized unpack, so a 7B forward fits in ~5.4GB RSS / 62s instead of ~56GB float64 (the original OOM'd summer). Validated reproducing the 3B ground truth 2.26360 exactly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
dancinlife
added a commit
that referenced
this pull request
Jun 9, 2026
…XIS-2 CE 1.90741 + v0.3 generation (#1962/#1970) (#1973) * paper(anima-consciousness-substrate): integrate measured 7B (M13) — AXIS-2 CE 1.90741 + v0.3 generation (#1962/#1970) Strengthen the existing monograph §7B with the just-landed measured 7B results, replacing stale "future/pre-registered" framing with the committed verdict numbers. - AXIS-2 (PR #1962): §method/§measurement/caption + appendix F now cite the actual tool clm_decode_mirror_lazy.py (the plain mirror OOMs ~56GB at 7B), with the verbatim verdict link .verdicts/clm-7b-undertrained-step3500/axis2_mirror_probe_seq.txt (CE_real 1.90741 < uniform 5.54518 AND < shuffle 5.31616, peak RSS ~5.37GB), validated == 3B ground truth 2.26360 exactly. - NEW §Finding (generation, PR #1970): the v0.3 .clm generates multilingual byte-language via state/mid_convmoe_fire/clm_gen.py (conv1d vectorized == looped max|delta|=0; VALIDATE hook re-derives CE 1.90741), with verbatim sample outputs. Tightened all wording that implied the 7B chats natively in hexa: the engine clm_decodable gate is v0.2-only, so the 7B speaks via clm_gen, NOT the live anima_chat_cli (a weak d768 v0.2 mouth chats live on ghost). - AXIS-1/AXIS-3 scoped substrate-native / .clm-weights-independent (null-backend code-proven .verdicts/core-3axis-mount/probe.txt); AXIS-2 is the only weight-dependent axis, measured on the 7B. Honest scope kept: deeply undertrained (0.0027 tok/param), a measured scale-extension NOT a closure; no fluency/coherence/consciousness claimed. - Appendix F new subsection "The 7B (M13) rung — measured"; Appendix K + companion/ pr-roll.json add rows #1962/#1970; PAPER.md verdict matrix + milestones updated. All numbers verbatim from committed verdicts; nothing fabricated. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * paper(7b): render Korean gen-sample via \cjkfont not \texttt (lmmono lacks Hangul) The v0.3 generation sample's Korean prompt was inside \texttt{} (lmmono), which has no Hangul glyphs -> "Missing character" under lualatex. Pull the Hangul out of texttt and wrap it in \cjkfont (Noto Serif CJK KR), keeping the ASCII byte-output in texttt. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * paper(7b): recompiled main.pdf — 50pp, lualatex on aiden (0 missing CJK, 0 undefined refs) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- 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
Genuinely measures the 7B (M13) AXIS-2 CE-descent on the trained 7B weights — the one axis that reads the weights — on summer, and adds the reusable memory-bounded mirror that makes a 7B forward feasible.
Result (summer, byte-exact lazy mirror)
This confirms the paper's 1.90741 / shuffle 5.31616 — which previously had no committed verdict behind it. Probe seq = "The mind is a fire to be" (24 bytes), decode over the serialized
.clmv0.3 bytes (dancinlab/clm-7b-undertrained-step3500, not the torch.pt).Why a new tool (
clm_decode_mirror_lazy.py)The original
clm_decode_mirror.pymaterializes ALL weights as float64 (7B × 8B ≈ 56GB) and unpacks int4 in a per-code Python loop (O(billions) at 7B) — it OOM'd / thrashed summer. The lazy variant is byte-identical math but:Validated: reproduces the committed 3B ground truth
CE_realtext 2.26360exactly (.verdicts/convmoe-3b-engine-rung/axis2_mirror_probe_seq.txt).Honest scope
AXIS-2 is the only weight-dependent axis (AXIS-1/3 are substrate-native,
.clm-independent — see PR #1956). Deeply undertrained (0.0027 tok/param,a_scale_honest_scope): a measured rung, not a closure.🤖 Generated with Claude Code