From a2c63ca7f9719ca38ef57b2f9d0bb251e016e678 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Wed, 29 Jul 2026 18:37:19 -0500 Subject: =?UTF-8?q?=E5=B0=8F=E9=98=B6=E6=A2=AF=E6=8E=92=E9=98=9F(=E7=94=A8?= =?UTF-8?q?=E6=88=B7=E8=AE=BE=E8=AE=A1:=E5=90=91=E4=B8=8Bscale=E5=8F=8D?= =?UTF-8?q?=E6=8E=A8=E5=90=91=E4=B8=8A):=20C256/C384=20FW=E6=96=B0?= =?UTF-8?q?=E6=A1=A3+C512=E6=B3=84=E6=BC=8F=E4=BB=AA=E5=99=A8(=E7=8E=B0?= =?UTF-8?q?=E6=88=90ckpt)=20=E2=80=94=20=E5=9B=9B=E7=82=B9=E7=94=BBgap(?= =?UTF-8?q?=E5=AE=BD=E5=BA=A6)+dose*(=E5=AE=BD=E5=BA=A6)=E5=8F=8C=E5=AE=9A?= =?UTF-8?q?=E5=BE=8B,=E5=88=A4=E5=88=AB=C3=971=E2=86=92=C3=97128=E6=98=AF?= =?UTF-8?q?=E5=B9=82=E5=BE=8B=E8=BF=98=E6=98=AF=E5=AE=BD=E5=BA=A6=E9=98=88?= =?UTF-8?q?=E5=80=BC;=20=E9=A1=BA=E5=B8=A6=E9=87=8F=E5=8C=9672M=20gap?= =?UTF-8?q?=E9=87=8C=E5=8F=AF=E5=85=B3=E9=97=AD=E4=BB=BD=E9=A2=9D;=20?= =?UTF-8?q?=E9=93=BE=E5=9C=A8dg128=E5=90=8E=E8=87=AA=E5=8A=A8=E6=8E=A5?= =?UTF-8?q?=E5=8A=9B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- docs/campaign/CASCADE_ABLATION_PLAN.md | 10 ++++++++++ 1 file changed, 10 insertions(+) (limited to 'docs') diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index f3678b3..8d0ed5c 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,16 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### QUEUED (2026-07-29, user design): WIDTH MINI-LADDER — scale DOWN to extrapolate UP. +New FineWeb rungs below the existing pair: C256 (~27M, 0.53B tok) + C384 (~46M, 0.93B tok), +L12/hd64, Chinchilla-matched, EP(bsign)+BP twins; plus the leak instrument AT C512 (existing +fw72m_bsign_s190000: dgain dose {1,4,16} + allbp bar). Four-point deliverables: +gap(width) and dose*(width). Key discriminator: C512->C768 dose jumped x1 -> x128 — smooth +power law (dose ~ width^k -> soft wall, budget k for 270M) vs width THRESHOLD near C600 +(leak absent below -> different physics, sharper story). Also quantifies whether the 72M +0.076 gap hides a small closable leak (lk512 arms answer directly). Chained behind fw135m_dg128 +on all three GPUs; ~12h of ladder after the validation completes. + ### RESULT 72 (2026-07-29): THE CURVE GOES ALL THE WAY — dg128 closes 97% (tail within 0.0015 of the all-BP bar); the physics leak is FULLY REDEEMABLE by displacement amplification. FULL-DOSE VALIDATION LAUNCHED. -- cgit v1.2.3