xuan/memory/MEMORY.md

• Qwen3-4B CPT Experiment — 20B token CPT on 8xH100, completed 2026-05-22, final ckpt at global_step_19074
• Environment Setup — conda env ttt, VeOmni, dependencies, lr_decay_ratio gotcha
• N4 Transfer Info — sigma@135.232.252.197:50000, SSH key not configured yet
• Qwen3-4B-NHS Model — YoungXuan/Qwen3-4B-NHS; In-Place-TTT inference silently DISABLES its TTT (zero conv)
• NIAH Eval Results — Single-needle NIAH saturated (Base 9.96, NHS 10.00), not useful for paper
• BABILong Eval Results — earlier NHS "boost" INVALID (TTT was a no-op); Base-v2 is the valid baseline
• TTT Inference Findings — faithful TTT replay collapses; NS conditions update DIRECTION; needs tuned-down length-aware lr
• Qwen3-0.6B TTT Experiment — In-Place-TTT 1k-step train; faithfully reproduces ttt_lr=3 grad explosion (in progress)
• Inference Always cf-patch — RULE: inference always uses cf-patch fast-weight update; report ONLY cf-patch score, never the "vanilla"/no-TTT column
• TTT Train Recipe Findings — RULE: ttt_norm_preserve=TRUE is mandatory (+20-30pt RULER, fixes loss-blowup); inner η insensitive plateau w/ norm=true; test η~3 (paper's value)