summaryrefslogtreecommitdiff
AgeCommit message (Expand)Author
13 daysPSGD-QUAD入列: vendor psgd.py + Optimizer子类适配器(合成损失closur...Yuren Hao
13 days种子恢复: rlin_s2精确续@80k(~48h), bp_s3续@260k(~14h, weights-only冷...Yuren Hao
13 days筛选正式闭卷(11优化器33臂): AdamW收口≈5.01; 增益钳位测试...Yuren Hao
13 daysRESULT 90终局: NS价值曲线k=3以60%成本买回99%(质量门槛远低于...Yuren Hao
13 daysRESULT 90: 优化器筛选收敛 — Muon 4.7227 领跑, NS消融判决(摘NS...Yuren Hao
13 days偏置放大公理(用户08-06指出): 逐坐标归一化把持久小偏置...Yuren Hao
13 days第二代原语优化器: SignLine/ConsensusLion/DitherLion 实装+本地筛...Yuren Hao
13 daysEqOLion R4(真实轨迹判决): 冻结极因子几十步退相关至~0.53=慢...Yuren Hao
13 days优化器价目表实装: Lion/OLion/Adafactor2D/build_alt + --opt 六选项Yuren Hao
13 daysEqOLion设计+当天三轮证伪: O(1)秩一探针死(两轮), 真实动量...Yuren Hao
13 days优化器审计修订: OLion(sign输出→正交化器精度要求塌到1bit)...Yuren Hao
13 days表征方案修订(用户纠正): 优化器建模义务不因不造板而消失Yuren Hao
13 days硬件线转向(08-05用户决定, Samuel建议): 元件表征方案取代MVP...Yuren Hao
2026-08-04HF上传: 135M EP/BP权重对(剥优化器状态542MB各)+tokenizer+notebook+...Yuren Hao
2026-08-04RESULT 89终局 + 交付物: 135M完赛 EP 3.2087 vs BP 3.2074(+0.1% ppl), ...Yuren Hao
2026-08-04RESULT 89: 135M判决 — 修复版3.2024 vs BP双seed均值3.2054(gap -0.003...Yuren Hao
2026-08-03预写: 给Ben信中的"请教K"段(用户08-03决定跑完135M后征询)Yuren Hao
2026-08-03RESULT 88: K电池判决 — K=1/2/3在CE上不可分(配对差3e-5, 比seed...Yuren Hao
2026-08-03RESULT 87: KHS论文精读 — 同性质获发明人背书, 但迭代计数...Yuren Hao
2026-08-03RESULT 86: K=1梯度余弦亦0.9999 — 记录对我方不利的证据Yuren Hao
2026-08-03RESULT 85: EP成本定案 — 受控实测wall-clock 2.70×(纯训练步2.77...Yuren Hao
2026-08-02审计补充: Rasmuskh=Rasmus Kjær Høier=KHS论文二作(本人核对commit...Yuren Hao
2026-08-02RESULT 84续: 135M第二进度点确认(rlin 3.3958 vs BP 3.3999, 旧读出+0...Yuren Hao
2026-08-02审计: EP/coupled-learning公开实现的对比读出地板普查(33仓库)Yuren Hao
2026-08-02工具: contrast_floor_check.py — 对比读出地板的独立自检探针Yuren Hao
2026-08-02技术摘要更新: 72M真数字3.400 vs 3.397(+0.4%), 135M改为进行中(...Yuren Hao
2026-08-02RESULT 84: 135M中途裁决 — 修复版3.4576 vs BP 3.4693(EP领先1.2%ppl)...Yuren Hao
2026-08-01RESULT 83补记: 135M吞吐0.70→2.05 it/s(GPU3被他项目占满;timan108 C...Yuren Hao
2026-08-01RESULT 83: 72M真gap=+0.0036(0.4%ppl,≈seed散布0.0029) — 修复回收0.1...Yuren Hao
2026-08-01技术摘要(给Eder/慈善出资方): 五尺寸对照表+读出地板结果+...Yuren Hao
2026-08-01RESULT 82: 五宽度阈值律闭合 — 最小位移比1.4e-5→4.2e-8单调...Yuren Hao
2026-08-01RESULT 81: 无操作证伪测试三尺寸全通过(Δ≤0.009<<seed散布), ...Yuren Hao
2026-07-31RESULT 80: 2×阶梯五点曲线完赛 — C128/C192/C320 EP与BP统计不可...Yuren Hao
2026-07-31Dillavou通话材料v3: 机制段按RESULT 77-79改判(读出地板+零成本...Yuren Hao
2026-07-30RESULT 79: CE终审+100%关闭(尾窗3.5206=allbp杆,四位小数命中预注...Yuren Hao
2026-07-30RESULT 78: read_lin梯度级验证==fp64(fp32下泄漏塌零,读出舍入=全...Yuren Hao
2026-07-30RESULT 77: 地板理论三箭全中(fp64塌零/β平移/β×8痊愈) — 物...Yuren Hao
2026-07-30RESULT 76: 谱仪读出 — 逐块泄漏剖面(C768上6块-0.70..-0.23,C512...Yuren Hao
2026-07-30RESULT 75: dg128完赛口径分裂(best回带+0.049/尾窗未回带+0.133,关...Yuren Hao
2026-07-30dgain几何profile + M1泄漏谱仪(用户设计,07-30)Yuren Hao
2026-07-29RESULT 74: Phase M发射计划(用户批准) — 2×减半阶梯(C128/C192/C3...Yuren Hao
2026-07-29RESULT 73: 治理条令(用户裁决) — 正向scaling claim暂停;三类严...Yuren Hao
2026-07-29小阶梯排队(用户设计:向下scale反推向上): C256/C384 FW新档+C51...Yuren Hao
2026-07-29RESULT 72: 剂量曲线通顶 — dg128关闭97%(距all-BP杆0.0015),物理...Yuren Hao
2026-07-29RESULT 71: 剂量曲线对数律(34/50/63%@×4/8/16),位移等效律三连验...Yuren Hao
2026-07-29RESULT 70: 首批真杠杆 — dg4top+34%(cos0.83下CE胜出),β收益=100%...Yuren Hao
2026-07-29RESULT 69: 饱和盲区被自己的治疗测试证伪(cap30/50关闭0.0000,...Yuren Hao
2026-07-28RESULT 68: 泄漏定位收官 — 100%住顶半block梯度,attn占2/3,与R44...Yuren Hao
2026-07-28RESULT 67: 仪器重标定(0.0685=33%harness confound+67%物理0.0462); 电...Yuren Hao
2026-07-28RESULT 66: head被节流到60%(位移微探针:全族1.00-1.02唯W_out 0.60) ...Yuren Hao