Record: 11-gram Eval Cache + Hedge Mixer (val_bpb: 0.8609)#909

Open

sunnypatneedi wants to merge 26 commits intoopenai:mainfrom

sunnypatneedi:submission/v10-moonshot-0.8609

Commits on Mar 24, 2026

v6.0 moonshot: LoRA TTT + In-Place TTT + surprise gating + eval-only XSA

sunnypatneedi
and
claude
committed
Add Full GPTQ (Hessian-aware quantization) — 31% quant gap reduction

sunnypatneedi
and
claude
committed
Fix critical bugs in In-Place TTT: add scoring + weight reset

sunnypatneedi
and
claude
committed
Update README with full v6.0 feature set and experiment plan

sunnypatneedi
and
claude
committed
Fix artifact size (disable int7) + reduce TTT epochs (20→5)

sunnypatneedi
and
claude
committed
Disable GPTQ — root cause of 0.18 bpb quant penalty in Run 1

sunnypatneedi
and
claude
committed
Run 0 (baseline) + Run 1 (LeakyReLU only) — retro-disciplined approach

sunnypatneedi
and
claude
committed
Run 0/1 rebased on merged SOTA PR #414 (1.1228), not unverified PR #548

sunnypatneedi
and
claude
committed
Run 2: + temperature calibration sweep (T=0.95-0.99)

sunnypatneedi
and
claude
committed
Run 3: int5 quantization + 3.5x MLP (~33.6M params in 16MB)

sunnypatneedi
and
claude
committed
Add submission template — ready to fill after 3-seed validation

sunnypatneedi
and
claude
committed
Session 3 retro: In-Place TTT falsified, GradQuant over budget

sunnypatneedi
and
claude
committed
Fix flash_attn_interface import + add SDPA fallback for all runs

sunnypatneedi
and
claude
committed
Fix flash_attn import: try FA3, then FA2, then SDPA fallback

sunnypatneedi
and
claude
committed
v8.0 Phase 1: PR #549 baseline + TTT enabled (2 lines changed)

sunnypatneedi
and
claude
committed

Commits on Mar 25, 2026

Commits on Mar 26, 2026

Commits on Mar 27, 2026