import arm qwen3/head_k16_solo_o0: anchors (every close), stats, results, log

import/qwen3/head_k16_solo_o0/closes/qimp_head_k16_solo_o0_close1_s15_qwen_import.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5652ac3a63a10130db1ce9003e129b64cae79ab41cbeb231938b98e2a848bdda
+size 54819504

import/qwen3/head_k16_solo_o0/closes/qimp_head_k16_solo_o0_close2_s15_qwen_import.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62fafccc111241140b1ef15ebdc44c0c47f2ddc9c1ba9a5a3de33254de19f048
+size 54819504

import/qwen3/head_k16_solo_o0/manifest.json

@@ -0,0 +1,18 @@
+{
+ "tag": "qimp_head_k16_solo_o0",
+ "run": "head_k16_solo_o0",
+ "convention": "qwen3",
+ "member": "s15_qwen_import",
+ "final_anchor": "qimp_head_k16_solo_o0_close2_s15_qwen_import.safetensors",
+ "closes": [
+  "qimp_head_k16_solo_o0_close1_s15_qwen_import.safetensors",
+  "qimp_head_k16_solo_o0_close2_s15_qwen_import.safetensors"
+ ],
+ "finished": true,
+ "form": "head",
+ "base": "solo",
+ "steps": 2000,
+ "trunk_step": 245674,
+ "shipped": "2026-10-08 18:16 local",
+ "note": "every close's anchor ships (mids included); the final close's anchor is the run's s15_qwen_import.safetensors"
+}

import/qwen3/head_k16_solo_o0/s15_qwen_import.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62fafccc111241140b1ef15ebdc44c0c47f2ddc9c1ba9a5a3de33254de19f048
+size 54819504

import/qwen3/head_k16_solo_o0/stats.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9054902ef0e9f030dcedbe288957bce372a3fb775bc025bdd0ccf20ac1214e7
+size 264325

import/qwen3/head_k16_solo_o0/train.log

@@ -0,0 +1,96 @@
+[    0.3s] THE JOB: qimp_head_k16_solo_o0: the Qwen import arm s15_qwen_import (head form; the teacher Qwen3 0.6B base at depth 16) alone on the bare trunk; blocks all; 2000 steps, a close every 1000; lambda 2.0 off-domain, 0.0 on the caption rows; pure Adam lr 0.001; the compiled chain with recompute; card 30.5 GB with 30.4 GB free; host RSS 0.6 GB
+Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
+[   17.2s] mini-beatrix-3 @ 245,674: trunk loaded from /workspace/qarm/dl/mini-beatrix-3/checkpoints/step_00245674.safetensors and frozen (376.0M); alephllm 0.10.7 at /workspace/qarm/alephllm/geolip/alephllm; amoe-lora 0.2.11; torch 2.8.0+cu128
+[   17.8s] tokenizer /workspace/qarm/dl/qwen3_tokenizer (gpt2 spelling; vocab 151669; sha256 be75606093db2094): 'a taco truck' -> ['a', 'Ġtaco', 'Ġtruck']; sites [(1, 6, 7)]
+[   27.2s] the teacher: Qwen3 0.6B base (/workspace/qarm/dl/qwen_3_06b_base.safetensors, sha256 cd2a512003e2f9f3; 28 layers, width 1024; config /workspace/qarm/dl/qwen3_06b); depth 16 = the input of layer 16
+[   27.2s] the read's captions: 512 of each draw of /workspace/qarm/dl/coco_caps_2x2048.json (4,090 captions of both draws kept out of training)
+[   40.3s] caption pack (/workspace/qarm/dl/mini-beatrix-3/caption_pack/pack_v2, train): dff 121,153 rows, scf 42,080 rows, danbooru 603,026 rows, cc12m 1,242,176 rows, coco 19,227 rows; host RSS 4.2 GB
+[   48.9s] read rows draw1: 512 captions (0 unspellable or too long), rows (512, 99), 4,920 sites; Qwen's states at ['16', '28post'] (fp32; final-state size 132.0)
+[   53.0s] read rows draw2: 512 captions (0 unspellable or too long), rows (512, 151), 5,086 sites; Qwen's states at ['16', '28post'] (fp32; final-state size 131.5)
+[   53.0s] streams: web text; partners none (two web rows a quiet chunk); held-out sets ['captions', 'web'] x 8 batches
+[arms] attached 's15_qwen_import' for phase 'qwen_import': 13.69M params x 32 blocks, lambda 2.0, spec {'n_slots': 16, 'K': 16, 'D': 8, 'hidden': 256}, recompute on, compile on
+[   64.6s] BIRTH: the member attached over the trunk: web rows read exactly as before; the hand-run trunk against the model's forward: armed 0.0e+00, masked 0.0e+00 (gates: 0.0)
+[   76.0s] standardization statistics from 13,911 reference sites on 8 batches (hers per block; Qwen's depth-16 states: size 61.7, sigma range 0.486..5.375); saved /workspace/qarm/out/anchors/qimp_head_k16_solo_o0_stats.pt
+[  214.7s] THE HEADS (linear, her standardized states -> Qwen's standardized depth-16 states, born as fp64 ridges, 139 s): held-out R2 per block 8: 0.351, 12: 0.397, 16: 0.432, 20: 0.458, 24: 0.438, 28: 0.385, 31: 0.334; 33.6M head parameters; saved /workspace/qarm/out/anchors/qimp_head_k16_solo_o0_frame.pt
+[  492.2s] CLOSE 0 (step 0): the import read on draw 1 (512 captions, 4,920 sites); per block: her reading against Qwen's final state, masked -> armed / against depth 16 masked -> armed; context (token means removed) masked -> armed
+[  492.2s]    block  8: final 0.426 -> 0.426 / depth 0.496 -> 0.496; context 0.285 -> 0.285; mse 0.866 cos 0.512
+[  492.2s]    block 12: final 0.470 -> 0.470 / depth 0.548 -> 0.548; context 0.331 -> 0.331; mse 0.827 cos 0.546
+[  492.2s]    block 16: final 0.508 -> 0.508 / depth 0.579 -> 0.579; context 0.372 -> 0.372; mse 0.785 cos 0.577
+[  492.2s]    block 18: final 0.514 -> 0.514 / depth 0.581 -> 0.581; context 0.386 -> 0.386; mse 0.762 cos 0.593
+[  492.2s]    block 20: final 0.526 -> 0.526 / depth 0.573 -> 0.573; context 0.399 -> 0.399; mse 0.759 cos 0.595
+[  492.2s]    block 24: final 0.562 -> 0.562 / depth 0.526 -> 0.526; context 0.466 -> 0.466; mse 0.792 cos 0.571
+[  492.2s]    block 28: final 0.540 -> 0.540 / depth 0.452 -> 0.452; context 0.448 -> 0.448; mse 0.852 cos 0.527
+[  492.2s]    block 31: final 0.475 -> 0.475 / depth 0.385 -> 0.385; context 0.385 -> 0.385; mse 0.900 cos 0.494
+[  492.2s]    draw 2 (the repeat), armed against the final state at 12/16/20/24/28: 0.471 / 0.508 / 0.529 / 0.567 / 0.546
+[  492.2s]    I1 the reach: the largest rise over her bare reading +0.000 at block 31 (draw 2 +0.000 at 31) against the bar +.10; I2 the context: +0.000 at block 31 against +.10; bits: captions +0.00000, web +0.00000 against +.012
+[  492.2s]    gate mean 0.047 / widest 0.047; compiled chain: 2 cache entries (limit 256), 0 limit warnings; docs drawn 256 (skipped {'test': 2, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 6.9 GB; host RSS 9.1 GB
+[  492.2s] THE INSTRUMENT CHECK: her bare reading against Qwen's final state on draw 1 (this driver / e038's row): block 8 0.426 / 0.426, block 12 0.470 / 0.470, block 16 0.508 / 0.509, block 20 0.526 / 0.526, block 24 0.562 / 0.562, block 28 0.540 / 0.540; largest difference 0.001
+[stream] opening 'fineweb-edu' (HuggingFaceFW/fineweb-edu) …
+[  498.9s] step 1/2000: task 0.6398 (mse 0.640 cos 0.603); quiet A 0.00000 off 0.00000; 1,727 sites; 6.73 s/step; the job: 8.3 min spent, about 224.3 min of steps left; card peak 16.1 GB
+[  587.4s] step 50/2000: task 0.6273 (mse 0.627 cos 0.610); quiet A 0.00000 off 0.00021; 1,733 sites; 1.90 s/step; the job: 9.8 min spent, about 61.9 min of steps left; card peak 16.5 GB
+[  677.1s] step 100/2000: task 0.6158 (mse 0.616 cos 0.619); quiet A 0.00000 off 0.00025; 1,740 sites; 1.85 s/step; the job: 11.3 min spent, about 58.5 min of steps left; card peak 16.5 GB
+[  767.7s] step 150/2000: task 0.6093 (mse 0.609 cos 0.624); quiet A 0.00000 off 0.00028; 1,738 sites; 1.84 s/step; the job: 12.8 min spent, about 56.6 min of steps left; card peak 16.5 GB
+[  858.3s] step 200/2000: task 0.6107 (mse 0.611 cos 0.623); quiet A 0.00000 off 0.00028; 1,717 sites; 1.83 s/step; the job: 14.3 min spent, about 54.9 min of steps left; card peak 16.5 GB
+[  948.3s] step 250/2000: task 0.6034 (mse 0.603 cos 0.628); quiet A 0.00000 off 0.00028; 1,721 sites; 1.82 s/step; the job: 15.8 min spent, about 53.2 min of steps left; card peak 16.5 GB
+[ 1038.8s] step 300/2000: task 0.6165 (mse 0.617 cos 0.621); quiet A 0.00000 off 0.00029; 1,750 sites; 1.82 s/step; the job: 17.3 min spent, about 51.6 min of steps left; card peak 16.5 GB
+[ 1129.6s] step 350/2000: task 0.6118 (mse 0.612 cos 0.624); quiet A 0.00000 off 0.00031; 1,725 sites; 1.82 s/step; the job: 18.8 min spent, about 50.1 min of steps left; card peak 16.5 GB
+[ 1220.5s] step 400/2000: task 0.5988 (mse 0.599 cos 0.632); quiet A 0.00000 off 0.00031; 1,772 sites; 1.82 s/step; the job: 20.3 min spent, about 48.6 min of steps left; card peak 16.5 GB
+[ 1311.1s] step 450/2000: task 0.6032 (mse 0.603 cos 0.629); quiet A 0.00000 off 0.00034; 1,827 sites; 1.82 s/step; the job: 21.9 min spent, about 47.0 min of steps left; card peak 16.5 GB
+[ 1401.7s] step 500/2000: task 0.5945 (mse 0.595 cos 0.634); quiet A 0.00000 off 0.00032; 1,767 sites; 1.82 s/step; the job: 23.4 min spent, about 45.5 min of steps left; card peak 16.5 GB
+[ 1492.6s] step 550/2000: task 0.6891 (mse 0.689 cos 0.633); quiet A 0.00000 off 0.00034; 1,682 sites; 1.82 s/step; the job: 24.9 min spent, about 44.0 min of steps left; card peak 16.5 GB
+[ 1583.0s] step 600/2000: task 0.6383 (mse 0.638 cos 0.636); quiet A 0.00000 off 0.00038; 1,717 sites; 1.82 s/step; the job: 26.4 min spent, about 42.4 min of steps left; card peak 16.5 GB
+[ 1673.9s] step 650/2000: task 0.6000 (mse 0.600 cos 0.633); quiet A 0.00000 off 0.00044; 1,805 sites; 1.82 s/step; the job: 27.9 min spent, about 40.9 min of steps left; card peak 16.5 GB
+[ 1765.1s] step 700/2000: task 0.6300 (mse 0.630 cos 0.639); quiet A 0.00000 off 0.00036; 1,672 sites; 1.82 s/step; the job: 29.4 min spent, about 39.4 min of steps left; card peak 16.5 GB
+[ 1855.7s] step 750/2000: task 0.6746 (mse 0.675 cos 0.639); quiet A 0.00000 off 0.00036; 1,708 sites; 1.82 s/step; the job: 30.9 min spent, about 37.9 min of steps left; card peak 16.5 GB
+[ 1946.7s] step 800/2000: task 0.5835 (mse 0.583 cos 0.643); quiet A 0.00000 off 0.00035; 1,771 sites; 1.82 s/step; the job: 32.4 min spent, about 36.4 min of steps left; card peak 16.5 GB
+[ 2037.4s] step 850/2000: task 0.5889 (mse 0.589 cos 0.640); quiet A 0.00000 off 0.00036; 1,665 sites; 1.82 s/step; the job: 34.0 min spent, about 34.8 min of steps left; card peak 16.5 GB
+[ 2128.3s] step 900/2000: task 0.5908 (mse 0.591 cos 0.641); quiet A 0.00000 off 0.00037; 1,848 sites; 1.82 s/step; the job: 35.5 min spent, about 33.3 min of steps left; card peak 16.5 GB
+[ 2219.8s] step 950/2000: task 0.6426 (mse 0.643 cos 0.637); quiet A 0.00000 off 0.00040; 1,781 sites; 1.82 s/step; the job: 37.0 min spent, about 31.8 min of steps left; card peak 16.5 GB
+[ 2311.0s] step 1000/2000: task 0.5880 (mse 0.588 cos 0.642); quiet A 0.00000 off 0.00036; 1,745 sites; 1.82 s/step; the job: 38.5 min spent, about 30.3 min of steps left; card peak 16.5 GB
+[ 2311.4s] anchor saved: /workspace/qarm/out/anchors/qimp_head_k16_solo_o0_close1_s15_qwen_import.safetensors
+[ 2453.2s] CLOSE 1 (step 1000): the import read on draw 1 (512 captions, 4,920 sites); per block: her reading against Qwen's final state, masked -> armed / against depth 16 masked -> armed; context (token means removed) masked -> armed
+[ 2453.2s]    block  8: final 0.426 -> 0.424 / depth 0.496 -> 0.498; context 0.285 -> 0.285; mse 0.829 cos 0.539
+[ 2453.2s]    block 12: final 0.470 -> 0.467 / depth 0.548 -> 0.554; context 0.331 -> 0.329; mse 0.782 cos 0.578
+[ 2453.2s]    block 16: final 0.508 -> 0.505 / depth 0.579 -> 0.594; context 0.372 -> 0.365; mse 0.733 cos 0.613
+[ 2453.2s]    block 18: final 0.514 -> 0.514 / depth 0.581 -> 0.601; context 0.386 -> 0.380; mse 0.706 cos 0.632
+[ 2453.2s]    block 20: final 0.526 -> 0.526 / depth 0.573 -> 0.599; context 0.399 -> 0.396; mse 0.696 cos 0.639
+[ 2453.2s]    block 24: final 0.562 -> 0.566 / depth 0.526 -> 0.574; context 0.466 -> 0.465; mse 0.707 cos 0.632
+[ 2453.2s]    block 28: final 0.540 -> 0.562 / depth 0.452 -> 0.520; context 0.448 -> 0.462; mse 0.743 cos 0.608
+[ 2453.2s]    block 31: final 0.475 -> 0.524 / depth 0.385 -> 0.475; context 0.385 -> 0.426; mse 0.774 cos 0.587
+[ 2453.2s]    draw 2 (the repeat), armed against the final state at 12/16/20/24/28: 0.470 / 0.504 / 0.528 / 0.568 / 0.566
+[ 2453.2s]    I1 the reach: the largest rise over her bare reading +0.049 at block 31 (draw 2 +0.050 at 31) against the bar +.10; I2 the context: +0.041 at block 31 against +.10; bits: captions +0.20747, web +0.00013 against +.012
+[ 2453.2s]    gate mean 0.197 / widest 0.264; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 32,523 (skipped {'test': 44, 'spelling': 0, 'empty': 0, 'short': 0}); card peak 16.5 GB; host RSS 10.9 GB
+[ 2543.8s] step 1050/2000: task 0.5734 (mse 0.573 cos 0.650); quiet A 0.00000 off 0.00035; 1,763 sites; 1.81 s/step; the job: 42.4 min spent, about 28.8 min of steps left; card peak 16.5 GB
+[ 2635.6s] step 1100/2000: task 0.5690 (mse 0.569 cos 0.653); quiet A 0.00000 off 0.00035; 1,731 sites; 1.82 s/step; the job: 43.9 min spent, about 27.3 min of steps left; card peak 16.5 GB
+[ 2726.0s] step 1150/2000: task 0.5654 (mse 0.565 cos 0.656); quiet A 0.00000 off 0.00036; 1,714 sites; 1.82 s/step; the job: 45.4 min spent, about 25.8 min of steps left; card peak 16.5 GB
+[ 2816.5s] step 1200/2000: task 0.5711 (mse 0.571 cos 0.653); quiet A 0.00000 off 0.00036; 1,724 sites; 1.82 s/step; the job: 46.9 min spent, about 24.2 min of steps left; card peak 16.5 GB
+[ 2907.3s] step 1250/2000: task 0.5731 (mse 0.573 cos 0.651); quiet A 0.00000 off 0.00037; 1,695 sites; 1.82 s/step; the job: 48.5 min spent, about 22.7 min of steps left; card peak 16.5 GB
+[ 2997.6s] step 1300/2000: task 0.5736 (mse 0.574 cos 0.651); quiet A 0.00000 off 0.00049; 1,701 sites; 1.81 s/step; the job: 50.0 min spent, about 21.2 min of steps left; card peak 16.5 GB
+[ 3088.9s] step 1350/2000: task 0.6183 (mse 0.618 cos 0.649); quiet A 0.00000 off 0.00045; 1,710 sites; 1.82 s/step; the job: 51.5 min spent, about 19.7 min of steps left; card peak 16.5 GB
+[ 3179.6s] step 1400/2000: task 0.5606 (mse 0.561 cos 0.659); quiet A 0.00000 off 0.00039; 1,767 sites; 1.82 s/step; the job: 53.0 min spent, about 18.2 min of steps left; card peak 16.5 GB
+[ 3270.3s] step 1450/2000: task 0.6227 (mse 0.623 cos 0.650); quiet A 0.00000 off 0.00096; 1,719 sites; 1.82 s/step; the job: 54.5 min spent, about 16.7 min of steps left; card peak 16.5 GB
+[ 3360.1s] step 1500/2000: task 0.5770 (mse 0.577 cos 0.649); quiet A 0.00000 off 0.00067; 1,670 sites; 1.81 s/step; the job: 56.0 min spent, about 15.1 min of steps left; card peak 16.5 GB
+[ 3451.1s] step 1550/2000: task 0.6081 (mse 0.608 cos 0.660); quiet A 0.00000 off 0.00047; 1,703 sites; 1.81 s/step; the job: 57.5 min spent, about 13.6 min of steps left; card peak 16.5 GB
+[ 3541.8s] step 1600/2000: task 0.6386 (mse 0.639 cos 0.660); quiet A 0.00000 off 0.00043; 1,705 sites; 1.81 s/step; the job: 59.0 min spent, about 12.1 min of steps left; card peak 16.5 GB
+[ 3633.2s] step 1650/2000: task 0.5997 (mse 0.600 cos 0.659); quiet A 0.00000 off 0.00044; 1,780 sites; 1.82 s/step; the job: 60.6 min spent, about 10.6 min of steps left; card peak 16.5 GB
+[ 3724.0s] step 1700/2000: task 0.5649 (mse 0.565 cos 0.657); quiet A 0.00000 off 0.00044; 1,733 sites; 1.82 s/step; the job: 62.1 min spent, about 9.1 min of steps left; card peak 16.5 GB
+[ 3814.9s] step 1750/2000: task 0.5623 (mse 0.562 cos 0.659); quiet A 0.00000 off 0.00042; 1,704 sites; 1.82 s/step; the job: 63.6 min spent, about 7.6 min of steps left; card peak 16.5 GB
+[ 3906.1s] step 1800/2000: task 0.5587 (mse 0.559 cos 0.663); quiet A 0.00000 off 0.00037; 1,777 sites; 1.82 s/step; the job: 65.1 min spent, about 6.1 min of steps left; card peak 16.5 GB
+[ 3996.9s] step 1850/2000: task 0.5556 (mse 0.556 cos 0.662); quiet A 0.00000 off 0.00046; 1,728 sites; 1.82 s/step; the job: 66.6 min spent, about 4.5 min of steps left; card peak 16.5 GB
+[ 4087.7s] step 1900/2000: task 0.6013 (mse 0.601 cos 0.661); quiet A 0.00000 off 0.00041; 1,836 sites; 1.82 s/step; the job: 68.1 min spent, about 3.0 min of steps left; card peak 16.5 GB
+[ 4178.6s] step 1950/2000: task 0.6373 (mse 0.637 cos 0.666); quiet A 0.00000 off 0.00042; 1,742 sites; 1.82 s/step; the job: 69.6 min spent, about 1.5 min of steps left; card peak 16.5 GB
+[ 4269.8s] step 2000/2000: task 0.5653 (mse 0.565 cos 0.657); quiet A 0.00000 off 0.00087; 1,747 sites; 1.82 s/step; the job: 71.2 min spent, about 0.0 min of steps left; card peak 16.5 GB
+[ 4270.3s] anchor saved: /workspace/qarm/out/anchors/qimp_head_k16_solo_o0_close2_s15_qwen_import.safetensors
+[ 4412.1s] CLOSE 2 (step 2000): the import read on draw 1 (512 captions, 4,920 sites); per block: her reading against Qwen's final state, masked -> armed / against depth 16 masked -> armed; context (token means removed) masked -> armed
+[ 4412.1s]    block  8: final 0.426 -> 0.424 / depth 0.496 -> 0.503; context 0.285 -> 0.288; mse 0.822 cos 0.544
+[ 4412.1s]    block 12: final 0.470 -> 0.461 / depth 0.548 -> 0.555; context 0.331 -> 0.327; mse 0.771 cos 0.584
+[ 4412.1s]    block 16: final 0.508 -> 0.494 / depth 0.579 -> 0.594; context 0.372 -> 0.359; mse 0.720 cos 0.621
+[ 4412.1s]    block 18: final 0.514 -> 0.500 / depth 0.581 -> 0.599; context 0.386 -> 0.372; mse 0.700 cos 0.635
+[ 4412.1s]    block 20: final 0.526 -> 0.509 / depth 0.573 -> 0.598; context 0.399 -> 0.389; mse 0.685 cos 0.645
+[ 4412.1s]    block 24: final 0.562 -> 0.536 / depth 0.526 -> 0.588; context 0.466 -> 0.437; mse 0.682 cos 0.647
+[ 4412.1s]    block 28: final 0.540 -> 0.539 / depth 0.452 -> 0.564; context 0.448 -> 0.437; mse 0.699 cos 0.635
+[ 4412.1s]    block 31: final 0.475 -> 0.522 / depth 0.385 -> 0.546; context 0.385 -> 0.419; mse 0.717 cos 0.624
+[ 4412.1s]    draw 2 (the repeat), armed against the final state at 12/16/20/24/28: 0.466 / 0.498 / 0.516 / 0.544 / 0.549
+[ 4412.1s]    I1 the reach: the largest rise over her bare reading +0.047 at block 31 (draw 2 +0.058 at 31) against the bar +.10; I2 the context: +0.034 at block 31 against +.10; bits: captions +1.30153, web +0.00025 against +.012
+[ 4412.1s]    gate mean 0.387 / widest 0.638; compiled chain: 5 cache entries (limit 256), 0 limit warnings; docs drawn 64,564 (skipped {'test': 97, 'spelling': 1, 'empty': 0, 'short': 0}); card peak 16.5 GB; host RSS 11.1 GB
+[ 4412.1s] DONE qimp_head_k16_solo_o0: 2000 steps in 60.6 min of steps (1.82 s/step); 2 closes; 4412 s in all; card peak 16.5 GB; host RSS 11.1 GB; compiled chain: 5 cache entries (limit 256), 0 limit warnings