Preserve item1 terminal results and freeze seed1 training review entry
This commit is contained in:
1 parent
888f706612
commit
c0f89a5e95
114 files changed
+282280
No files matched your search
@@ -631,3 +631,10 @@ Track A执行配方明确(Claude1794975093.432709):采用A同5split全实
|
||||
|
||||
|
||||
Track A训练入口轻审/启动:Assist1794975952实际PASS1fa923f8,执行方逐字固定b7879eec并调用原validate;不提前做postfit全审。2026-11-18 12:28唯一1262364服务实际派发1262432/1262422到GPU2/3,真实epoch1进度,三idle polls均PASS;源6701e4d6/合同be4ac1f6不变。item1继续、GPU0/1保留,首epoch成本仍待实测;只checkpoint冻结不自动test。启动事实不构成成本门或结果PASS。
|
||||
|
||||
|
||||
## Item1终态与符号订正(2026-11-18)
|
||||
|
||||
item1唯一恢复COMPLETE,Claude1794983276采用原分类UNIDENTIFIED_OR_OVERFIT_SIGNS,但其初次把负g解释成更多updates降低误差。Melt1794983333已立即用冻结statistics与实际端点纠正:g=error(U)−error(2U),负g是2U误差更高;s以log2实际嵌套组件数比归一化,正s才数据收益。原结果/判据不变,不重算;不将此结果当作长程训练获益、更多数据无效或过拟合的因果证明。Track D数据瓶颈触发未满足;新长程候选只能纸面且不能建立在错误符号上,后续解释须由Claude订正。Track A继续原训练,postfit实际新PASS983bba99待执行方采用,五训练全终态前不得预测。seed1按2860/3024原五fold/seed1/200epochs,118800总GPU秒及动态闲卡分发准备轻审;不重训seed0、不跑seed2/ensemble或自动扩展科学范围。
|
||||
|
||||
Claude1794983369已明确撤回初次g方向误读并订正设计记录3c05a46:负g表示2U更高误差;数据validation信号不等于test确认,原分类不变。长程候选请求撤回,仅将未来纸面方向改为过拟合控制,仍在seed1后按阶段机制讨论,不授权新训练。Melt报告摘要的s使用log2实际嵌套组件数比,g明确U−2U;原raw结果保持。
|
||||
@@ -1,3 +1,13 @@
|
||||
## 最新 seed1入口已固定待轻审;Track A postfit PASS已采用(2026-11-18)
|
||||
|
||||
seed1根/home/frank14f/MeltMind-runs/phase3-seed1-2026-11-18,implementationd2f552d1aa54a107b935c9614ded2c5d7157f693/contract4036fd7a58412c775a57a7a8fadf727b9b440b812c2188062556c12bcb735dd8/proofd9a80bf680727669d8e5ca05c68eb99ca89704b79a2d39c28c8da5baaf2f3ba7,38codes;原S2科学epoch AST只shuffle种子/等待hook不同,15CPU检查PASS,旧fixture失败保留。按Claude2860/3024,动态next-pending→首个已释放GPU,GPU2给Track A postfit优先;GPU0/1先行、3待原训练退出、2待原评估退出。118800GPU总/111600训练/22320每fold+7200postfit,phase3原item账72421.315999一次;无review/started/训练,须本次轻审。报告docs/phase3-seed1-entry-2026-11-18.md。
|
||||
|
||||
Track A实际pre-score PASS983bba99已逐字接收、freeze并validate,绑定b164071e/19faa4a9/83e178ef/8f0f3c10/30pins保持;root review-adoption记录,无预测/目标访问/等待器。原1262364训练尚未全终态,不提前启动postfit。item1终态结果已交Claude;其初稿负g方向误读已在1794983333指出,报告摘要明确g=U−2U,负值为2U更差,不据此宣称长训练收益。
|
||||
|
||||
## 最新 item1 COMPLETE,禁止重启(2026-11-18)
|
||||
|
||||
唯一恢复1341741已exit0,1..6全部PASS、评分PASS;旧686hash复用,4582预测固定后552unique targets/121test读一次,S1指标121复用。postfit GPU1229.244214/7200,CPU1167.436863/3600,训练71192.071785253一次,item总72421.315999/120000。原STOP保留。验证/外部复用test均UNIDENTIFIED_OR_OVERFIT_SIGNS;不是数据无效或过拟合因果证明。报告docs/phase3-item1-results-2026-11-18.md,root terminal-audit及原始score完整保留。下一步Claude阶段讨论;Track A原训练继续,postfit b164071e待审,seed1按2860/3024准备独立动态调度轻审入口,尚未启动。下方原运行快照为历史。
|
||||
|
||||
## 最新 Track A postfit 绑定已固定待独立审阅(2026-11-18)
|
||||
|
||||
Claude5970/2020/2202已实读:同原5split/1537test visits/586unique,复用A不重拟合;全5训练终态及退出后,GPU2串行预测、完整预测冻结再单次读目标。part内平均visits后unique part等权bootstrap2000/seed2026111408/95%percentile,无gate,保留family leakage。根/home/frank14f/MeltMind-runs/track-a-postfit-2026-11-18,implementationb164071e391c92ad8ee99dfa96c37f5b17991731/contract19faa4a9e14a19cedfeac3cbc5334462e10892f49269384d655449b525f6068e/proof83e178efae084bd39994a22b21d2d020a9cb46f144ba11dbc6a4e260f1b46124,comparison8f0f3c10,30源码固定。15CPU检查PASS;无review/started/预测/目标访问。45GPUh总账含7200postfit/3600CPU,旧训练23pins保持。详见docs/track-a-postfit-entry-2026-11-18.md;训练PASS不替代新pre-score PASS,不自动waiter/启动。
|
||||
|
||||
@@ -0,0 +1,30 @@
|
||||
# Item1 data-sufficiency postfit — terminal result, 2026-11-18
|
||||
|
||||
The single authorized recovery completed with all six resumed prediction guards and scoring PASS. Original predict0 was reused by hash (686 files), and the original zero-worker GPU-busy STOP remains unchanged. All4582 canonical predictions were frozen before552 unique target reads (121 test). S1 original121 metric rows were reused. No training or scientific source changed.
|
||||
|
||||
**Sign convention:** `s_2U = [error(50%,2U) − error(100%,2U)] / log2(K100/K50)`, where K is the actual nested learning-component count; positive means benefit from the larger pool. `g_50 = error(50%,U) − error(50%,2U)` and `g_100 = error(100%,U) − error(100%,2U)`; positive means benefit from more updates, negative means **higher error at2U**. Test50% error rises0.068848977→0.072841268; test100% rises0.066848549→0.069586623; validation100% rises0.058409554→0.061272961. This result does not support the statement that longer training reduced error.
|
||||
|
||||
Both validation and the reused external test fold retain the frozen classification `UNIDENTIFIED_OR_OVERFIT_SIGNS`. This does not establish that extra data cannot help, or prove overfitting as the cause. External test contrasts are descriptive on the reused fold; uncertainty covers evaluation components, not training seed/pool uncertainty. Update-budget contrasts jointly change the complete schedule and update count.
|
||||
|
||||
| Domain | Contrast | Value | 95% interval |
|
||||
|---|---|---:|---|
|
||||
| common_validation | s_U | 0.003906463 | [0.000680483, 0.007145781] |
|
||||
| common_validation | s_2U | 0.003269779 | [0.001118877, 0.005473240] |
|
||||
| common_validation | g_50 | -0.002212699 | [-0.006578530, 0.002078290] |
|
||||
| common_validation | g_100 | -0.002863407 | [-0.005320848, -0.000478482] |
|
||||
| common_validation | s_2U_minus_s_U | -0.000636684 | [-0.004435444, 0.003029046] |
|
||||
| common_validation | epoch_s_25to50 | 0.001920445 | [-0.002090456, 0.006225782] |
|
||||
| common_validation | epoch_s_50to100 | 0.002614435 | [-0.001558464, 0.006148817] |
|
||||
| common_test | s_U | 0.001957316 | [-0.001526956, 0.005417185] |
|
||||
| common_test | s_2U | 0.003184503 | [-0.000897271, 0.007680015] |
|
||||
| common_test | g_50 | -0.003992291 | [-0.007872280, -0.000289727] |
|
||||
| common_test | g_100 | -0.002738074 | [-0.005973915, 0.000581431] |
|
||||
| common_test | s_2U_minus_s_U | 0.001227187 | [-0.004228780, 0.006427824] |
|
||||
| common_test | epoch_s_25to50 | 0.005589158 | [0.001609502, 0.009629431] |
|
||||
| common_test | epoch_s_50to100 | 0.002279737 | [-0.000646325, 0.005217090] |
|
||||
|
||||
These are the original frozen relative-L2 contrasts, not new posthoc calculations. Positive s denotes the registered data-pool benefit; positive g denotes benefit from the longer update budget. Full arm/domain means, last-location and Kendall metrics, exclusions and all rows remain in `score/result.json`. Negative predictions are preserved without clipping.
|
||||
|
||||
Final cumulative postfit GPU 1229.244214134/7200s; CPU 1167.436862818/3600s. Original training71192.071785253s counted once; item total 72421.315999387/120000 GPU-s. All costs include the earlier STOP once.
|
||||
|
||||
Root: `/home/frank14f/MeltMind-runs/phase3-item1-postfit-recovery-2026-11-18`. Terminal audit and immutable source/model/prediction/review/terminal receipts are retained. Service exited0 and no item1 postfit process remains. No restart or rescoring. Stage interpretation and next-stage decision are submitted to Claude; Track A continues its already approved training, and seed1 remains separate reviewed-entry preparation.
|
||||
@@ -0,0 +1,15 @@
|
||||
# Phase3 seed1 training entry — 2026-11-18
|
||||
|
||||
Claude1794982860 approves118800 GPU-s total (111600 training,22320 per fold;7200 postfit), inside phase3's612000. Claude1794983024 specifies dynamic assignment of the next pending fold to the first allowed free GPU. This is the original seed-sensitivity step after item1, not a new recipe selected from its result. Original Mold specification is preserved byte-for-byte at references/2026-11-18/phase3-seed-sensitivity/mold-seed-ensemble-spec.md, SHA5dce5d0feb2b5cdd2703f53f0030b8127915d3464c7fa37e16507303b9c52b0d. Its seed2/ensemble portions remain future work.
|
||||
|
||||
New root `/home/frank14f/MeltMind-runs/phase3-seed1-2026-11-18`. Five fresh seed1 fits on original grouped folds0–4, same learning/validation/test roles,19 columns, full-node geometry, clocks, learning-only normalization, optimizer, schedule and200epochs. Validation every10epochs remains diagnostic only. Last-epoch checkpoint only; no seed0 refit, seed2, ensemble or target-based selection. Test predictions/scoring require a separate postfit implementation and pre-score review after all five terminal/model freezes.
|
||||
|
||||
Worker derives from immutable S2 worker. Scientific epoch-loop AST is identical after normalizing only the shuffle-seed literal and the first-epoch wait hook. Initialization seeds Python/NumPy/Torch/allCUDA to1; epoch shuffle is SeedSequence([1,epoch]); checkpoint records seed1. The original S2 helper disallowed fold0 and waited for all four folds at once, so the new helper admits exactly original folds0–4 and releases each fold after its own cost gate.
|
||||
|
||||
GPU0 is initially available. GPU1 becomes eligible after item1 recovery terminal and processes exit. GPU3 becomes eligible after all Track A training ends and exits. GPU2 is held for the higher-priority Track A postfit until that process has also terminated. Each next pending fold gets the first eligible idle card; actual assignment is written once, hashed and passed unchanged through guard to worker. There is no cross-card retry or reassignment of an attempted fold. All source pins and old failures are retained. Device idle polls and per-GPU locks remain required. A busy unrelated GPU causes a recorded STOP, not preemption.
|
||||
|
||||
Each running fold reprojects after every full epoch:1.25 times remaining training plus remaining every10epoch validation, using the greater of measured validation and its original same-fold maximum, with setup/elapsed costs and terminal allowance. Unmeasured folds keep their full22320s reservation; terminal actual costs replace it. This is a hard resource gate: the25% margin can stop a run even when an unpadded estimate is below the cap. No silent cap or margin relaxation. Track A has its separate45GPUh budget. Phase3's item1 settled72421.31599938683 GPU-s is included once; no previous S2 training cost is charged as new phase3 work.
|
||||
|
||||
Per fold15GiB VRAM/16GiB RSS/20GiB output; at most4 workers, aggregate64GiB/100GiB. One-time startup must check actual available memory/disk; when later GPUs become eligible, aggregate guards remain active. Limits are sampled, so overshoot/failures remain evidence. Training entry freezes models and stops; it never launches predictions or another stage.
|
||||
|
||||
Fifteen CPU synthetic tests PASS, including original five-fold roles/learning normalization, rejected test access, unchanged scientific epoch AST, all RNG and checkpoint seeds, fold-local cost barrier, full pending reserves/once-only prior ledger, validation costs, overrun STOP, dynamic dispatch and Track A GPU2 priority, hash-bound assignments, clean guard and rejection of monitor errors/recorded excess peaks. First test attempt failed because fixtures omitted parent directories before atomic writes (3 failures/2 errors); original fixture and receipt are retained, production science did not run. Metadata preparation read no scientific arrays, weights or targets. Exact light training-entry review is pending; test PASS does not substitute for independent review or launch evidence.
|
||||
Reference in new issue
Block a user