research: gobig_v2 C — per-step diagnostics 619727/619728 (frozen, run) #140

Open
starsetbyte wants to merge 2 commits from research/gobig2-c-diag-20261018 into research/gobig2-b-valmargin-20261018
Owner

gobig_v2 study C (diag) — 0-update per-step diagnostics, seeds 619727 (sharp-max) / 619728 (soft) vs peers

All 20 #133 checkpoints (gated_sharp_max and gated_soft, seeds 619719-619728) on PL16S = PL16 graphs 0-11 (7,546 views, N=4, lengths 0-16). From the model's own states, per step: prefix-alive rate, first-error class (op_swap = moved along the other operation's edge, stay, other), sign-aligned target state vs best non-target state (separation), readout gap, fraction of states outside the training state range; plus gate multipliers 2*softmax and relation projection. Parameters asserted unchanged.

  • Freeze GOBIG2-DIAG-FROZEN-001 06e37648…e1fd. Calibration 5.03 s / 4.99 s / 333 MiB; diagnose 22.40 s / 22.34 s / 333 MiB. All exit 0.
  • 619727 sharp-max: weakest gate selectivity of the 10 sharp seeds (A row 1.24/0.76; peers 1.30-1.48); separation median falls 1.42 -> 0.59 -> 0.00 (steps 1/8/16) while good peers hold ~1.8-2.0; 34% of states leave the trained range by step 16 (peers 0-14%); 84% of first errors are op_swap. Reading: the gate leaks the wrong channel and the error compounds with length.
  • 619728 soft: prefix-alive 0.995 at step 16; separation p05 stays 1.70-1.73 at steps 8-16, 0% out of range, only 4 first errors. 7 of 9 soft peers collapse (separation median ~0 by step 8). Its gate (1.56/0.44) and relation projection are not outliers, so no single parameter explains the extrapolation; it is the dynamics.
  • Study B (#139) shows the same 619727 checkpoint reaches PL16 depth 16 after 30 more updates, so its failure was under-training, not capacity.

See diag/diag-001/SUMMARY.txt and per-checkpoint JSON. Base: B branch (#139). Do not merge.

## gobig_v2 study C (diag) — 0-update per-step diagnostics, seeds 619727 (sharp-max) / 619728 (soft) vs peers All 20 #133 checkpoints (gated_sharp_max and gated_soft, seeds 619719-619728) on PL16S = PL16 graphs 0-11 (7,546 views, N=4, lengths 0-16). From the model's own states, per step: prefix-alive rate, first-error class (op_swap = moved along the other operation's edge, stay, other), sign-aligned target state vs best non-target state (separation), readout gap, fraction of states outside the training state range; plus gate multipliers 2*softmax and relation projection. Parameters asserted unchanged. - Freeze GOBIG2-DIAG-FROZEN-001 `06e37648…e1fd`. Calibration 5.03 s / 4.99 s / 333 MiB; diagnose 22.40 s / 22.34 s / 333 MiB. All exit 0. - 619727 sharp-max: weakest gate selectivity of the 10 sharp seeds (A row 1.24/0.76; peers 1.30-1.48); separation median falls 1.42 -> 0.59 -> 0.00 (steps 1/8/16) while good peers hold ~1.8-2.0; 34% of states leave the trained range by step 16 (peers 0-14%); 84% of first errors are op_swap. Reading: the gate leaks the wrong channel and the error compounds with length. - 619728 soft: prefix-alive 0.995 at step 16; separation p05 stays 1.70-1.73 at steps 8-16, 0% out of range, only 4 first errors. 7 of 9 soft peers collapse (separation median ~0 by step 8). Its gate (1.56/0.44) and relation projection are not outliers, so no single parameter explains the extrapolation; it is the dynamics. - Study B (#139) shows the same 619727 checkpoint reaches PL16 depth 16 after 30 more updates, so its failure was under-training, not capacity. See `diag/diag-001/SUMMARY.txt` and per-checkpoint JSON. Base: B branch (#139). Do not merge.
gobig_v2/diag: 0-update per-step diagnostics on PL16S
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 7s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 8s
d17c6207fc
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 7s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 8s
This pull request can be merged automatically.
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin research/gobig2-c-diag-20261018:research/gobig2-c-diag-20261018
git switch research/gobig2-c-diag-20261018

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch research/gobig2-b-valmargin-20261018
git merge --no-ff research/gobig2-c-diag-20261018
git switch research/gobig2-c-diag-20261018
git rebase research/gobig2-b-valmargin-20261018
git switch research/gobig2-b-valmargin-20261018
git merge --ff-only research/gobig2-c-diag-20261018
git switch research/gobig2-c-diag-20261018
git rebase research/gobig2-b-valmargin-20261018
git switch research/gobig2-b-valmargin-20261018
git merge --no-ff research/gobig2-c-diag-20261018
git switch research/gobig2-b-valmargin-20261018
git merge --squash research/gobig2-c-diag-20261018
git switch research/gobig2-b-valmargin-20261018
git merge --ff-only research/gobig2-c-diag-20261018
git switch research/gobig2-b-valmargin-20261018
git merge research/gobig2-c-diag-20261018
git push origin research/gobig2-b-valmargin-20261018
Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany!140
No description provided.