Record: GDN-Hybrid + Sliding Window Attention + compressed-code warmdown1000 - val_bpb 1.01671 (3-seed mean)

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/README.md

@@ -0,0 +1,29 @@
+# GDN-Hybrid + Sliding Window Attention + compressed-code warmdown1000 (cold-cache 3-seed mean 1.01671233 BPB)
+
+- **3-seed mean:** **1.01671233 BPB**
+- **3-seed std:** **0.00134386 BPB**
+- **Best seed:** **1.015700 BPB** (`seed 1337`)
+- **Worst seed:** `1.018237 BPB` (`seed 2024`)
+- **Artifact size range:** `15,713,422` to `15,903,365` bytes
+
+## Per-seed authoritative results
+
+| Seed | Steps | EMA BPB | Quantized BPB | XSA BPB | Artifact bytes |
+|------|------:|--------:|--------------:|--------:|---------------:|
+| 42 | 2227 | 1.007164 | 1.016200 | 1.021202 | 15,733,879 |
+| 1337 | 2242 | 1.007164 | 1.015700 | 1.020105 | 15,903,365 |
+| 2024 | 2227 | 1.009032 | 1.018237 | 1.024111 | 15,713,422 |
+| **Mean** | — | **1.007787** | **1.01671233** | **1.021806** | **15,783,555.33** |
+| **Std (sample)** | — | — | **0.00134386** | — | — |
+
+## Technique stack
+
+1. **SP1024 tokenizer** with a GDN-hybrid backbone (`[GDN×5] → SWA → [GDN×5] → SWA_shared`).
+2. **Fixed-predictor / no-TTT Track-A lane** — no eval-time or pre-quant adaptation in the scored artifact.
+3. **MuonEq-R + AdamW** training mix, EMA `0.997`, late QAT threshold `0.15`, and **warmdown=1000**.
+4. **GPTQ int6 + zstd-22** packaging.
+5. **Compressed-code record packaging** for `train_gpt.py`, `architectures.py`, and `configs.py`, which recovered artifact-size headroom without changing the trained model family.
+6. **Sliding-window attention side path** present in-model, but submission authority remains the pulled `quantized_bpb` values above.
+
+This record uses a fixed int6 model with **no TTT, no SLOT, no RLS, and no eval-time adaptation**. All three serialized artifacts are below the 16 MB cap. XSA telemetry is reported for completeness, but the submission authority remains `quantized_bpb`.
+

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/architectures.py

@@ -0,0 +1,5 @@
+import base64, lzma, linecache
+globals()['__file__'] = 'architectures.py'
+_src = lzma.decompress(base64.b85decode(b'{Wp48S^xk9=GL@E0stWa8~^|S5YJf5;A7bs%Uu8wXbs4O{*J2g^P4X}8A*?nKeo}D<QO>((oAtmlzXSM1Ig2`L`+cv@?w-@3r2bpI{8yz9`;@SXxb8{$fRtA3$`M_7l-P&$A)VpsI<{M1(k?)Sp%fgGHf{OA(e1{%%h>U(XyoMXIPp+maZ<`Y6aE!)~96}-?&KS_k=U~h?SsP>3(#hSXx;Gb}nY(jTA3=m<)>eOpgmMpv(N6Jw4N1{;sYZhPMhU0&O&Da!-;hJL0RkdivHzjmv^(5uZsfYeq27zGmIe`KoOHIQuYUjFMPp8l;u48XK7Wszs%jwMD3Ag^3BFZA|PV)yGfF&9VH-(s2Q|eiU=*2gP#b(g1_keckp+RTd1ebYAlkI_Q(1KdgIzo^2Ajexh6Vo<-BmT64<v4-1-B@Zcy2g-N}82Q<Ea*=hw3AZ>)g?{FY-8_?)s$@TwD&jnwEE94uXTfL)k!^9qGM3)Q`34S(SfW9B%*GWL@QNjM=A7Q9M@<u(gRT%zP@#PE!V{C+UBo+1IHl|$&CKbB<9~V0pkE~35`x<jBHXC}G)3sBisMOXu$^%bGFSi`<f~Bd4%sn={P=e{86L<@A<n--1wB}LOJz25ZlC_=nzccg{Rl}=S)3M6&X1NoN>9S#*t9C)TqR=33=ksZTjaD+X4M-pW80`pt#r_W8U_;|k+Ua{=p6)56RyNUVjjaM=XPL{;3+Y<$iDw6ztH0YE_r!|+t|eW=X<2P0w{@zd5iBFY$CTrD*n+ol+(E42zm}zhsa#B?qtO=*Lc{nP&hfDdkRtQ=Ep3rLq>hVos*{y!iUw=w2kYqLrMkW!EgT-IHvop$I@64dqxYqCV@_2s5AjSTf;6;^g}%+S<b`Qlt)i$@u^$(%=<Fn^$~BO+_|0snc?HA~$-Ti$=^!Tfm~Jpf1xjpCNVbO#!Oq?1Dt<LPzo-LQgJ(5gaj6CQwPbR{*wHUXA6$Oyt`%y#MgUM#!UXz)%o#5@p~zjNN;Jrc;n2+g=%K2-n=5zHKFi4zka>Zf;6p2bFm~5gVdaZ;l6u*JAf0kul^?N2uz0!sG3i*0K+|_D7}U1uB^#T(Yj%9S7V277Q}3e^KEP|TnWsSNT!n98yW)qh?F`Dne{<Yd1xy3SY#uYTFg`Gj+SGdkCCBPoH0p&1O@jA7plGXa)Mq+MjT+dxJqr!g(!@l`OV>&h-HQW_;jl)yQ(JD0+=SXB6lPhIl5y&+;)mb@J7se?c1nHe2f;xjtWiwr%87?ToT05m1D7Ht^7wfyHwC&oUu|2#kt3@McM!NN*Xc(8l3poPj324}xWH7HqS;NT18Ws9RV~4Jb>@o!@qB$3Rod_D-2^PR;B@BW*;pxUq=9KFG^`q&c=d>!=yylYsps!c<pZN}!NhMnN{~GJIrND%t-DTGz)wKBgH4AJY*7d^FyqK~@{a~puQzF;@vMm_nCih6o=v8~gI3otWAl%rWCqUSJ6);n)Ef9N0C|DI4tWZc^)Lw7eHWzO#k_VR3VF}2DvvASmgC<&p`7N=oYe#$d$yl{ynjYr4-5HF`DW8nA-h4Oj1GLGQGH5=dV($~R5)q;33CtFcM1>DqUp&i52VmLj6d^p4h8b^NT6U4z9CbVh$=ki+|kz+>sEkqyPVAkwo(WKMju~!K33A^GmP{d5wpd}cC<c0BF>>rGEH^_v|sZC>pk-|TC&?x5C&1Rrf_c^uk4hePX(VYWmHYq2yxenyjs>Yfo1y@K2Xm;pgPofy-WD7x^L`xCmBLInjBy|B5rtRG7`-X_RQx_vA!=wCi5)p_NE$)u#5le=Kip=y)gOL{>W!g584gaFlfPn2Mm`OuckQXPoC`4ykJQ?t?T~;IPnr@c6(Y0>Xi(9;rSI3(mK55h2yOusT&))QSD}Xr$z@0jT(LP!Z8jfo5%LAlprOO7;9m>d9})Qqc99#5Y$8$n!Oca<Hw2wMR8E-=mr`fW{T_@y8%GS<#1XT8NzM0$g~1tg{9rg%C){FzX(U6X#auj+D$l5jV0Ss-4@%z;)(@gYv6`-<G}F>g9SJ|J(&AkT9tR3=Ep8tYdo=BP$nEgDSOrG2*Uq%bHKitAx#+Ox%d78EPqpwjakDR+Dr!tnkELJ>vP9y;{XLy;%eyP9P@bC*dDiovwN|2OvUL5|ICP)ik0h;e=q<7iKiW2(+=ipjVsA%Pm5lk8}(T+hMK!GrqWo6mWTqJjgnh;h8&w3Z6jO=ORx!i`I#8ku6W*#@azgGJjEFo@ZeoRuq$0W!%8xJ?=Kr=H~3nLT`CZF(}?vt+0&L@w3j`2UpupFq;As;(s%VL1G}Y!UFNn#ueXiv06o}s21CoicTpwuuVeOz75vlEsW%_<QuA+H8P?GACiUkRgMGO|+Ww<*>(K#<Lx3aF=6eqf;3gQ}<ArSb{ItG|J=C!^+q*t@z@9N%AIb?)XEh~26$_$w=Ez+X+!)7r?|0(W11JJQbYgp{wX;2p`1+<a6PiG4Ls3vFit1qyRwLR>f5hlGL#P1P0#0h|z(EZidY3Z8dMji`PH&efpmyR-o?}(jL`3gmDksLIN$l}HL_7S*h~v{e6xsM3a_zc&5~tYyPn|q0(A{fJx}h4x{?J+=Kzj`#?i5q!K6NZy1OIDyReIZr^o8JPPlV6hY@fJB8vbE<uqRaaJsmhpP6|#JOX%v4D2(Wf(shGg?MvQ=6d~&e*{G<8y|(1M#g!t5n>&JuS72MPoTMbu{uoW)YXTJy_`|NXcDT1mgP%}3oj$nc?vPe-m|cPH5DvMkj=T`32U1H?b&!Cai5c@^>mFTtU(q;h_Srv2GN+>N-!`S-qe`6yhU`4v&-`Zf1vmg6DOleRZ~=7PiYtq0xqkDAs1^CmavP3W=1rYCO^60AD17J<L1>NN*DO@eqx3EQ`meU=>bQ+}A^PKmF3)1d@zn1FqVE*3t2MZ>lDV(JO&LEMhl)<l32)#M#1RNJ)C>?7qmOO%lI&Byxq`w__IVlA+oaI4yAP~CY1%&f`4TTAxcT6#bGA?~utOF<)!z7&Uc5J8o&=;`I0_c%&O7lxa`;KyQBtc=v;6(+C~@~I-0h)Pklxj2{t+K*0b^vtA%r|VB(kekV4P=(+)bGWe5QIezx}!<l1b0*-wzK1=FCT~!2tVA@ymFeM5?7X$C&)cI{1Mc`Zb5kmI5bN3kYEo;2)W!yPdA$zuRWE3KcS=sx%~u!%F@ng!{-FQ`QD(4mA$qsus5B9=cGp{;b6Yw~%2JW{ByNvPmG2-5)2OY?mEW;J8xTI@joSgNEPnFu80XHXG3+Ruqz&wO5iL=j}1g?$C0lHVBSXcj(wl!9C$_wKO+d__}*yp%o7CYJkx=!`H&iX54R9WZ<Pp{FPkFtQ{BB=F=-yd)C5%4CxJ|)#?7LEAKjzcSYqhpT~wSzVS9Gwc@%~8J^h~iK5ZxkBw+Tn?h@zscii#g?$D1g<rN<s%5;fFA9L#Vmh1eK5sdGEI0F;P(?V$u@x2lr}QxTI~SB~q8uP4S@W>&X}#)nCr|F>@$j+G<rRiFM|u!;tBkA>n&w86a8m4=i_-z``GsN9()Fg#o2F|P*D)ivEjUV>Ii^x)CY)P!er`IvAsr1PSZrqUwq*?>JMP-aa(VP?Y`}EA0{pD_VYM>t&J4lqU#?>MC-ZE94FVL<UIn3rr>#)WRp}ym5u$zku}HFYK^f|e8*?;gE$fQ$g-=xai0%n#7!{L^7gyoh_RQ-uSD6Zr`O8?{`Bs>&WZEo2&+eRHw2LpjTu+#ef^s_DBpj&NUua7j=EgZH(bR945C$%@ZZLQ^c-gF0bMI_BtQxdZC>LFk-5F^;S3~Lq6MdhzPLYzlUvzGWJJKMFeq1Ne7!$^F+-(Fpeyhelf_G6bHf8;ZGVidY_ko-k8M@de+KHwM3|h})k)Y>0rzcDk%{k?+2Om?kA`nFUp5ydxZ3HUfwvftwoBG}<oty@%e%lva9h!-%pz4e;1g(h(1msuu;VxLq21iM*9eHDHWkAo6wpPCxRb6yGBp;MR#KYxw(s?JC^$@df3cEI%5y=cBh731+Zir^v?DT}*Ve(F{RKg$9f=^Xsz5TZ_%4hrhDvlx6v24c{#4`CLpuRALavL;Xoh&XZ+5Rl6TdpM4y8ONI!~Rv^)2^r0VWdSdnv1_-;$3pdnB$p2{hUs70TaP>Dk0$%IJEGyMm@zZm~kF|5PF;9`zDGw7gLuzP82|o0yyo`ZdH>sL3u2P&o7}?C2C^PZ3myM-DPK)MS^7ikg0XYSiiNc1DMuW8$WCb*7Iy^>jLH$!uP3v5*g*eN=L2B^7pl9@53$H&cEV5i`ev*&$3ZCZi2<gGa1Th!a8>zhED2vUfz<p_qol0K$%a1DJ%Ltvte)4^}D2LG(7IWip+DHZ1{U-*2uNm&Z(6jcAq|R&VSIAFLrdY%kCdG5EMYlwN+{yZ9JQ$1{P!~n>7f`VwL+_`2+UD0LAy^9r;IcJp|FF<R(;JEM9+`-S=ldW>H2_YRp42GchCW1&r>Llk7as4FQh#y~L;lp20-;&-ZSzR*4Nt@MU~O5hK(+@9rQ?Ofq`;i)w<&D;(t-%agZq-U3vMviP}g<pm#*oHTqULUUq60Xhy71;cYy%X!e0C9@Ud*Srx!i3;5)-FbC3vmCq6g(AeH`DByozcER*b`$1&ed7Qh>Icwv0ry_`#uSdQa()%F&Jy%ptAx}+@iu)78qNMfI(?JEO*f5)O-n>(VW&5IjzA*lWs(S>6uZ6TNOYKN11|YqS)(ljRR94GHeG+uJ#ipZp3!F3nI9S@&0Gxu3d(YkjaPn#O4sJUC4@~eG{3e%kMierjR2g_`htusm|mB5)Y?)NRB|>XYVr$4od!rGOhzb6HXk$Fe4-jnr=?AU#HNpv8lqu0KpeomI8)I!PU*_}Eqo1&q3Em@EqlI>MPuI$R7<?B5j)1u*yJ?aQlc<$DJ7386vN6rkVjfE=VMcK=8M9J88EqzsMBTUS&gUOhh+96&b9;_Wqv8&IzL4<<~fAPz>Q+D2pCybW{OtorvR;iBdUlx1MkRW{}FHsjWaS*(-GE973A#%V9?Yc7$fNmwpIzh7;7cbO3eVQNitYRK{}WiK3SPBZ>PRZP?i>OvK#oCGXWcm_&w?4xFI9>P-&+H*XbbRJ9k3r)&YxqN9*_OI=3JS9JgRy*CZ5C&7>Ews3w-?#Ea@1iy}1bEZ`+4vmPSXPoc|lejLHIoU%sL-w$ygnC_j6<<n)-wnN{R!gD=nK2btfMoQ$d8{(35!I;*)i#>DjfUvISBKCOYpjiYgd&995_C?GN6Tj=MXumdE5pq;`PK^FiVa9Q2b?>*e-MtQdhH-xZA`VjHtAZ?Pt@h!(YS3dvQu5{RF_8?i(*qKG;qz0H;1LWnQMd=^6x|yTDQ~|X0DkYu9DJfZO@AR&Gb^Y$o-t{50rkVm!Y@uz>f8EqSmB2np~O?XMG_+ZMfKa2SY@gBv(FynMU>}>%1uFDlhkXvHMlNP$*x+iS_;Xve#O;2d3aS3wPO29oe08K^Ykwt&1H=ERDD0`tc2sQf&m+B=_=DO|4m7TX+9lC96R2FE-L*0fnx?-Yf)>xbdtyp@o?8WsT!Jq@#D#s8Qr2U;pU>~21-#ZvNj+h7D1Sbqp!S`=%xcJ=Os-b;Pew)*Cp>J2-16oU!O;DR43+)p~a=k>+o>%p-CFfl}-RVA75zTo`{rqu@y{M_E*GbenUK0kRh@Kp%<-9-y60jZ5HOD25_4GD4VJ`%Rr~TIb1M@8^lLdE`8p1LCW-LH}ZV(B==Tsh`MNvfIM`{$NU?47L$t*QjoP2>LCt+d#1qTPVCCClPGMc9Xf?xWeG?64q(0tV^SEQX<vB)twN4sVp}B3P{OOV<Qgc>yRdKp@(>@x3=2n)HVdb*;ie%yr1H!y%vxF9jaP^8RUz9eCG@_b9{?rKWpQ{<W)Qze^=XHO2S;IX0E?6?T)>BL3?J0l`29|+<>~K*M1f6(vNaQN8rt?%u$T7BSZX)^TNcDPvRves2A%<fY`@JjTqAb8Q#$fUQw1Q~nb|c6ptd9E`dE-QUXtXz_4XAY86>%MpG0(Sr=Gv4j5_t50%ibaVnFY{*L5~!ZHYS#-}c<Rx_IR1k@c#GX6c@)Am*R0Z7K8Wa;r!Zr;k67S=93_w~)qx6-X4KB@94!R!rDSwyl%#=w~=a8bqlgYbKh3_9YZ4SOCLj=nLyz1}~czhtDE<b;mEPO%)vqvYa=Rs#qFmKY~)gfyttDcc`~=hW&&S8-8PRgq|qg_cf1jLFFyG`Cw&6-$V}yKg1qTgvW;l{&;jCL`W&I<rmAw_|ocF#UVmgba=UNn`z@QlJuOU&2&meSjdnCgsm$S6tDMhue29@p-J>9N7Jyv2+%rQU|QbiWnQodW6%sie}?-af1QRSSL2lC8mVgVj3|}%9j2SWE-UJA>GNU0AN2Fw3ifb(bL8d{T!^))d^Y)H5AkbE(EMiY$)x&GQ6Fpisw<+OKXb9`jj;>`P#l3#%Rb}g<ClJ0`%(0vTpQkj;)RMW9O}f7fLRP7Qz#9iHeKu@$$WYbTQes(pGET4_D`kiGl$6Tq*dc!&~-axVfq^XlX~s(I1;4ArdA*Pg`C4px^mH_Z%M~r;tNG2RS?jLqzq5G;Gku<0+h*^&K+QKh{Di<XU1@{+5_S4Q%PCRIEF1^`RIQPlj?D%-of1g&+Mp&(FwA<k6bbM@hR^RI7L#|<Q@gTSQ_5lp9*QbX?DE41d|X6F!KfybTpa(0}au__NJ5l8B{><u9SwridQDmraT+}YMQ6A-5y?q4AV}!FShrPOXHo9_kZYeN+-Zyt?w|G#JLyzJgCc0*jz~4x!zysE+F|cUo0i6XH#{eIfK`V5mnOFx%=DRT$%aV6?m=OTjSp1Wsjqy{0(gZV1%rJW!K-)dFa9rBEY#=y|P2(ZGILJ8vn!u08qkpR3{2O$T;n<LENL`^9Gs3{o`CUj%8QKx4DhWy^N3s9RhwQgUo9O>Lk21)TkQrqFCfO7)__6_EQiI+LeqisY={74y)gr+Z0OWyXJ|O`3v*~^^#A3&HNYKi2_#&l)GuH(e$VkJqE~bRpE6D8pj)1)Pa<JTBMg<49e`0RNtp@Jhj9$fWdaC2FJRx_GKiEg+W86emX=yb*~#V)xwXnFU)#SB4;8zA$}aUzuIlw!#;)NfR}v+)j7EKF?~`(G#x-wfrF=Ko=pObG!AJ3=B$1do+>3IA)${Bb9W;KBJ)pDVVL7{cdGE*T>C_PE}ZNHWrUbarV+~pL=7j!>NZxVqBVT(ICrMygQ-N8G;v*LAj0e>5Mhjq#(Z`Z;&fxC_9TQz=x<$7AOogW^`A7jJW;_r5IqZ$Bb|>@v~B&7&NGAQJ4O<;?uCxMSV=qU-Ism)5zjl`o^=yLf>(64MGO^z9M~$?N|q;%{j1wxL+6ctZ)TK40ELOxc^V|EGyEa=CqXK9IFaHo0cR&X)e<Lu1m36Wl;56UD7o=bt`6Cjq0r7rx&eo_wr2%+tVfSa{`{2<hhw0~_X4{<ScaO3v$EHM#UdJ$z}cRS{bkrMg}6sb*j3$9`Q*N|bGanaGu~-E1O==`R%5%EW3R^lFaeSl#pSuR`~b>QI+?Osg~f9j(P7w=>`hhB2oLhAwba|b^l$6Zw{BB3D%pICv~{&e^*@?yX%`kvM*}lFXjXF`QXN`V%%Lbi5gECrFIu$^j*rjGdh6c33!Lnx0H156gF$<(UCSY8hnwRqTu+?Q(Yq2{RF_@!|Cfhz`6nO321;32A40s#PAQid&SP&s=*eQ2cuA}g8ZFv;cP)S)yKG-wFkwU#LP614f$)#|>(^5-@fr-vlPHx&J@S}O8Q&gSdvrEOD?~m-oDbtouDTi?zxg||zS{xX#D@l3fs{L+IaFLTS7OEswe)16cOv!C0S-=R0c-51+&1Q;i7mT}4?&S3bFa(M8X-rgpf63<&Qa%5?XSs@RXSBPYB&La_icC`VZDlfF>rP?=<BU%sR&g4w(M&zgu`*!y|00<j)_f{7)~rc+}aJ!0s_R>TAq|bXq!vl_}3^Sdt_RTm|e4Jqe*@p^8H*fh-p$FYdl5_a%6$faS5bcqrL3XLP{O#VFkV&)A6g~_cTjOAOGor+Rs^v0Wpui4Aav$W_SPq!EW3W;+CKj00HMQ+Q$I^sD?+7vBYQl0ssI200dcD')).decode()
+linecache.cache['architectures.py'] = (len(_src), None, _src.splitlines(True), 'architectures.py')
+exec(compile(_src, 'architectures.py', 'exec'))

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/configs.py

@@ -0,0 +1,5 @@
+import base64, lzma, linecache
+globals()['__file__'] = 'configs.py'
+_src = lzma.decompress(base64.b85decode(b'{Wp48S^xk9=GL@E0stWa8~^|S5YJf5;1K==%Uu8wXb(;1LpQ-~@JH;7bq<!2rM>sfvNj2i-6-~M*R0t23rwIc7?+`^!VD6uagcoK6DbS{9T~420xAWik!$biWb)rrE+C-SGfkiwhV0~r&>A7MzvaxD&2QbJ{2wIdK)(s%&mtZ-@C^j&Il;>(HV9Kdzl26iSpQ@}+W8F+gFd4uyfywfyMFM`tZ6yPk2@}Bf`F{`G|QqY_~SLO)^$j(lVw5|@~26gr#@Cy$?g#DNu)(MZ3zM3D848h-pG^4;H!1seE8;!)*!vr<*hh_M)n_J*FkAv3&y+~EaG*xEwZ(y0{Dluk1)M?vK|<?M15&9sh9@)sm=9EQ!l9<F=3hI{V_#FZb(VhKFRJTCvj$idas@TZXBNDLU;Z*-1PC|^dt;z;sGN1$)2AK@%4k~y7Ij08HFB(E_tT8(BIU{1VYBaM8e3);f(C<O!9w?`>r-z=2q;$f?Np>RAmtd0SnvDX9u|NdpKRv1XCu)-hRaPhu(+3vIVgK=%?W`h<nrj@^ClP57G3H$Ka_p0^Lhhy){F)Rzt1mZ4=&&wEnHFK4UVYgugE;)jv{i-?&ldyTBS(mqu&f5X^|fs%L$&s&;0h<_JOdpj5J4B-WfP{dfC9+{^sg^v5;>*Ir!o`XM-gyZHpY(;A=WQ%uIy-LEpk?32}<@uwZHb{=xO>@YgZg&n)w-2CzKB!f!PX4{sbfc`i@s;=Z)Js>=-iM6SYI+Q(`U`M<IVh8XMFDd+B`sqIQM5;%)xR;J+D72uYc+Iz<mDBMb+;NowsU><Q-aeFf6J}m)nODlO?Mv84+2(r)&&4T&86TFCk=yXWPs~%Vy-LioO;Yd>gV{}<wkto7{`3JW|8UjLBhjgROZzy&cg`aRWW%}9i{#DRsIpaD9_Va<3?@d{W`OB1T5`_6m`oCMDbS3M7BzqM<WQX(Nvd|X!atM(ztWo!P$;vfz$#P<5xGfqqg7;Yr-so}n=-sgRPRsFA9Fl0F%Pb@36LC1gv#~HAKY1Ml^jurbyw~$C$RHK^VlPtr!MLXq3wTKOCFZ_P3-79YWYpaCz_BWcE_YNptPThG8=ZVmwW2lGk<RbZgtH|@~c!AI7~_|56dJYhXhj_=hL<YLZz{c3mGo}L36^tMzj=fJx_u*k6!31n!ByTnf?<Me-7elw4W;X!EBx6Cs~oZ(X}TztV4Z82EEqhG0{8x{QBoG9kwpH*|JMc$m8jCmaz7f5-J%nY(HgqSIzgomvEiNbsB~$Tk@6G8mHWLI=%U?eAy?aPo6mX!x#wD74cjcK>Z-mMs{`4u_n2wsGC14_Pi&1O3GbJ82-<=%ePMG6(KXaL$5W57RMv}0_vMs`@UkDBTH-R%ko})Yg?AOCht_=5RD#cvY2)jQ&tE*`7?1>pJTzY%sLiRkLAqYJr0qyGdhd~uy@0D5xG=3C?waVq~$|Q#!<?>JCX&-oh{sufJ>q}>z-_UbAeQ3l;2VRB$MCD-_NXhRZ#7v%p&MFs!t`_Y$`GiU4-YOmqjOJm$ZOflTGRdW}mWmDGGLnOW%_=;rSv3Gm$hgn<~9@hs)#-g{3+d;Vo~}ax{{l%l=VDe2Bo^B8Lcu<75bM-hIE)J8?E}g2~HGoyh5%fYjL46`UGK$i-9Ta^wXSK-7bJn%)z#XM+i(MFA#)^J?TaE?@kIZ=V4zfp$|^j3BuWKRuaz*OmfluO>S6^XY2?e9Z3q2+1CVv*C!s#1PtU@V=DVxEFsS>8lv-tc~<X(o7~cC62?Guo8WU2W#K{m_c%|lavV8mW;Tmc+d*b>n;l5Q+cxVMb98Ou7vHP5Xf53+_dNGNVPzcF-5<Y@s`%l@sGT&_6tnrf2ac(0YTLqkxs8D|5jppc#LkSCX%PAd`v&<7r1YYXy}Hg@^(PC)Kz<^CrSVSBK&Om4$c`m00HL<{~-VX(Bw{7vBYQl0ssI200dcD')).decode()
+linecache.cache['configs.py'] = (len(_src), None, _src.splitlines(True), 'configs.py')
+exec(compile(_src, 'configs.py', 'exec'))

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/requirements.txt

@@ -0,0 +1,5 @@
+flash-linear-attention
+zstandard
+sentencepiece
+# flash_attn_3 is pre-installed in runpod/parameter-golf:latest
+# torch, numpy, and other standard deps are pre-installed in the image

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/submission.json

@@ -0,0 +1,40 @@
+{
+  "name": "GDN-Hybrid + Sliding Window Attention + compressed-code warmdown1000",
+  "val_bpb": 1.01671233,
+  "bytes_total": 15903365,
+  "blurb": "SP1024 GDN-Hybrid fixed-predictor run with warmdown=1000 and compressed-code packaging. Three-seed mean 1.01671233 across seeds 42/1337/2024, with all artifacts under the 16,000,000-byte cap.",
+  "author": "Joshua Martinez",
+  "github_id": "joshkmartinez",
+  "date": "2026-04-12",
+  "base_pr": 1545,
+  "val_bpb_method": "quantized_bpb",
+  "seeds": {
+    "42": {
+      "val_bpb": 1.016200,
+      "bytes_total": 15733879,
+      "steps": 2227
+    },
+    "1337": {
+      "val_bpb": 1.015700,
+      "bytes_total": 15903365,
+      "steps": 2242
+    },
+    "2024": {
+      "val_bpb": 1.018237,
+      "bytes_total": 15713422,
+      "steps": 2227
+    }
+  },
+  "mean_bpb": 1.01671233,
+  "std_bpb": 0.00134386,
+  "architecture": "SP1024 GDN-Hybrid ([GDN×5] -> SWA -> [GDN×5] -> SWA_shared), MuonEq-R + AdamW, EMA 0.997, warmdown=1000, GPTQ int6 + zstd-22, compressed-code packaging.",
+  "compliance": {
+    "artifact_under_16mb": true,
+    "training_under_600s": true,
+    "fixed_predictor": true,
+    "no_ttt": true,
+    "no_slot": true,
+    "no_rls": true,
+    "three_seeds": true
+  }
+}

records/track_10min_16mb/2026-04-12_GDN_Hybrid_DeltaRule_Warmdown1000_Compressed/train_seed1337.log

@@ -0,0 +1,90 @@
+W0412 21:30:56.460000 842497 torch/distributed/run.py:851] 
+W0412 21:30:56.460000 842497 torch/distributed/run.py:851] *****************************************
+W0412 21:30:56.460000 842497 torch/distributed/run.py:851] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0412 21:30:56.460000 842497 torch/distributed/run.py:851] *****************************************
+=== Direction-5: GDN Hybrid Training ===
+Arch: D_GDN_Hybrid (ARCH_MODE=D)
+Seed: 1337, Max steps: 9999, Warmdown: 1000
+Train seq_len: 2048, Wallclock budget: 590.0s
+QK_GAIN_INIT: 5.0
+World size: 8, Grad accum: 1
+EMA decay: 0.997, SWA: True (every 50)
+Late QAT threshold: 0.15
+MuonEq-R: enabled
+Validation tokens: 62,021,632
+Model built in 0.2s
+Parameters: {'embedding': 925697, 'recurrent': 13251360, 'attention': 792584, 'mlp': 18880512, 'other': 12800, 'total': 33862953}
+Total params: 33,862,953
+Matrix params: 33,263,616
+Scalar params: 75,049
+Embed params: 524,288
+Generated coprime shard order: stride across 80 shards
+
+================================================================================
+Starting training: max 9999 steps (from step 0)
+Wallclock budget: 590.0s
+================================================================================
+
+step     1/9999 | loss 6.9316 | lr_mul 0.0500 | mom 0.850 | 0.38 steps/s | 3s
+step     2/9999 | loss 6.7037 | lr_mul 0.1000 | mom 0.850 | 0.68 steps/s | 3s
+step     3/9999 | loss 6.1494 | lr_mul 0.1500 | mom 0.851 | 0.94 steps/s | 3s
+step     4/9999 | loss 5.8672 | lr_mul 0.2000 | mom 0.851 | 1.16 steps/s | 3s
+step     5/9999 | loss 5.8456 | lr_mul 0.2500 | mom 0.851 | 1.35 steps/s | 4s
+step     6/9999 | loss 5.7912 | lr_mul 0.3000 | mom 0.851 | 1.51 steps/s | 4s
+step     7/9999 | loss 5.7134 | lr_mul 0.3500 | mom 0.851 | 1.66 steps/s | 4s
+step     8/9999 | loss 5.7339 | lr_mul 0.4000 | mom 0.852 | 1.78 steps/s | 4s
+step     9/9999 | loss 5.6268 | lr_mul 0.4500 | mom 0.852 | 1.90 steps/s | 5s
+step    10/9999 | loss 5.5281 | lr_mul 0.5000 | mom 0.852 | 2.00 steps/s | 5s
+step   100/9999 | loss 3.6791 | lr_mul 1.0000 | mom 0.870 | 3.51 steps/s | 29s
+step   200/9999 | loss 2.7052 | lr_mul 1.0000 | mom 0.890 | 3.65 steps/s | 55s
+step   300/9999 | loss 2.4796 | lr_mul 1.0000 | mom 0.910 | 3.69 steps/s | 81s
+step   400/9999 | loss 2.3839 | lr_mul 1.0000 | mom 0.930 | 3.72 steps/s | 108s
+step   500/9999 | loss 2.3145 | lr_mul 1.0000 | mom 0.950 | 3.74 steps/s | 134s
+step   600/9999 | loss 2.2768 | lr_mul 1.0000 | mom 0.950 | 3.75 steps/s | 160s
+step   700/9999 | loss 2.2446 | lr_mul 1.0000 | mom 0.950 | 3.76 steps/s | 186s
+step   800/9999 | loss 2.2263 | lr_mul 1.0000 | mom 0.950 | 3.77 steps/s | 212s
+step   900/9999 | loss 2.2322 | lr_mul 1.0000 | mom 0.950 | 3.77 steps/s | 239s
+step  1000/9999 | loss 2.1975 | lr_mul 1.0000 | mom 0.950 | 3.78 steps/s | 265s
+step  1100/9999 | loss 2.1804 | lr_mul 1.0000 | mom 0.950 | 3.78 steps/s | 291s
+step  1200/9999 | loss 2.1805 | lr_mul 1.0000 | mom 0.950 | 3.78 steps/s | 317s
+step  1300/9999 | loss 2.1583 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 343s
+step  1400/9999 | loss 2.1301 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 370s
+step  1500/9999 | loss 2.1270 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 396s
+step  1600/9999 | loss 2.1350 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 422s
+step  1700/9999 | loss 2.1260 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 448s
+step  1800/9999 | loss 2.1136 | lr_mul 1.0000 | mom 0.950 | 3.79 steps/s | 474s
+step  1900/9999 | loss 2.1167 | lr_mul 1.0000 | mom 0.950 | 3.80 steps/s | 500s
+step  2000/9999 | loss 2.1099 | lr_mul 1.0000 | mom 0.950 | 3.80 steps/s | 527s
+step  2100/9999 | loss 2.0972 | lr_mul 1.0000 | mom 0.950 | 3.80 steps/s | 553s
+step  2200/9999 | loss 2.1001 | lr_mul 1.0000 | mom 0.950 | 3.80 steps/s | 579s
+Wallclock limit reached (590s), will stop after this step
+Stopping early at step 2242 (wallclock limit)
+
+Training complete in 590s (2242 steps)
+Peak memory: 35748 MiB
+Steps/sec: 3.80
+
+=== Applying EMA weights ===
+EMA BPB (no XSA): 1.007164
+Saved raw EMA model
+
+=== GPTQ: generating autoregressive calibration data ===
+GPTQ: generated 64 sequences, collecting hessians...
+GPTQ: collected hessians for 29 layers
+
+=== Quantizing to int6 + zstd-22 ===
+Artifact: 15,903,365 bytes (15.17 MB)
+
+=== Roundtrip Validation (quantized model) ===
+Quantized BPB (no XSA): 1.015700
+Quantization degradation: +0.008536
+Quantized BPB (XSA-all): 1.020105
+
+================================================================================
+FINAL RESULTS — D_GDN_Hybrid seed=1337
+  Training: 2242 steps, 590s
+  EMA BPB:       1.007164
+  Quantized BPB: 1.015700
+  XSA BPB:       1.020105
+  Artifact:      /root/pg-repo/records/track_10min_16mb/2026-04-12_JM_GDN_Hybrid_Warmdown1000_CompressedCode_Seeds42_1337_2024/checkpoints/final_model_D_GDN_Hybrid_seed1337.int6.ptz
+================================================================================