Datasets:
Download logs/run.log from syvb/qwen3-last-position-patching: direct link, hf CLI and curl.
- Browser
- Download file 91.3 kB
-
https://huggingface.co/datasets/syvb/qwen3-last-position-patching/resolve/main/logs/run.log
- Command line
-
hf download hf://datasets/syvb/qwen3-last-position-patching/logs/run.log
-
curl -L -o run.log https://huggingface.co/datasets/syvb/qwen3-last-position-patching/resolve/main/logs/run.log
91.3 kB
== 22:12:47 install
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning.
== 22:13:10 download data
[90mHint: The `hf-cli` skill is not installed. Run `hf skills add -g` to teach your AI agents how to use the `hf` CLI.[0m
Fetching 10 files: 0%| | 0/10 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 10 files: 20%|██ | 2/10 [00:00<00:00, 14.55it/s]Fetching 10 files: 50%|█████ | 5/10 [00:00<00:00, 20.71it/s]Fetching 10 files: 80%|████████ | 8/10 [00:00<00:00, 17.68it/s][32m✓ Downloaded[0m
path: /workspace/act-experiments/data/dl/sample/10BT/000_00000.parquet
== 22:13:19 test
Fetching 10 files: 100%|██████████| 10/10 [00:10<00:00, 1.69s/it]Fetching 10 files: 100%|██████████| 10/10 [00:10<00:00, 1.08s/it]
Fetching 10 files: 0%| | 0/10 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 10 files: 10%|█ | 1/10 [00:00<00:01, 6.55it/s]Fetching 10 files: 40%|████ | 4/10 [00:00<00:00, 17.86it/s]Fetching 10 files: 70%|███████ | 7/10 [00:00<00:00, 21.51it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 19%|█▊ | 58/310 [00:00<00:00, 576.47it/s]Loading weights: 37%|███▋ | 116/310 [00:00<00:00, 564.78it/s]Loading weights: 56%|█████▌ | 173/310 [00:00<00:00, 404.70it/s]Loading weights: 70%|███████ | 218/310 [00:00<00:00, 318.40it/s]Loading weights: 82%|████████▏ | 254/310 [00:00<00:00, 320.11it/s]Loading weights: 93%|█████████▎| 289/310 [00:00<00:00, 262.24it/s]Loading weights: 100%|██████████| 310/310 [00:01<00:00, 309.90it/s]
patch site -1: max relative error resid 1.7e-06
patch site -1: max relative error attn 2.9e-06
patch site -1: max relative error mlp 3.2e-06
patch site -1: max relative error logits 1.3e-06
patch site 0: max relative error resid 2.4e-06
patch site 0: max relative error attn 3.9e-06
patch site 0: max relative error mlp 3.9e-06
patch site 0: max relative error logits 4.2e-06
patch site 1: max relative error resid 2.2e-06
patch site 1: max relative error attn 4.3e-06
patch site 1: max relative error mlp 3.0e-06
patch site 1: max relative error logits 1.4e-06
patch site 14: max relative error resid 5.3e-06
patch site 14: max relative error attn 7.8e-06
patch site 14: max relative error mlp 8.6e-06
patch site 14: max relative error logits 6.0e-06
patch site 28: max relative error resid 1.7e-06
patch site 28: max relative error attn 2.9e-06
patch site 28: max relative error mlp 3.2e-06
patch site 28: max relative error logits 3.4e-07
OK
[W1014 22:13:31.519030231 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:13:31 means 0.6B
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 100%|██████████| 310/310 [00:00<00:00, 8214.72it/s]
Fetching 10 files: 100%|██████████| 10/10 [00:14<00:00, 2.02s/it]Fetching 10 files: 100%|██████████| 10/10 [00:14<00:00, 1.42s/it]
Fetching 13 files: 0%| | 0/13 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 13 files: 15%|█▌ | 2/13 [00:00<00:00, 13.06it/s]Fetching 13 files: 31%|███ | 4/13 [00:00<00:00, 15.96it/s]Fetching 13 files: 46%|████▌ | 6/13 [00:00<00:00, 16.99it/s]Fetching 13 files: 77%|███████▋ | 10/13 [00:00<00:00, 19.50it/s]Fetching 13 files: 92%|█████████▏| 12/13 [00:13<00:01, 1.93s/it]Fetching 13 files: 100%|██████████| 13/13 [00:22<00:00, 3.05s/it]Fetching 13 files: 100%|██████████| 13/13 [00:22<00:00, 1.71s/it]
Fetching 14 files: 0%| | 0/14 [00:00<?, ?it/s]Fetching 14 files: 7%|▋ | 1/14 [00:00<00:01, 9.99it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 14 files: 29%|██▊ | 4/14 [00:00<00:00, 14.95it/s]Fetching 14 files: 50%|█████ | 7/14 [00:00<00:00, 19.99it/s]34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
Fetching 14 files: 64%|██████▍ | 9/14 [00:20<00:00, 19.99it/s]6591/34019 docs
Fetching 14 files: 71%|███████▏ | 10/14 [00:25<00:14, 3.69s/it]10533/34019 docs
13510/34019 docs
Fetching 14 files: 79%|███████▊ | 11/14 [00:38<00:15, 5.25s/it]15912/34019 docs
17940/34019 docs
19703/34019 docs
Fetching 14 files: 86%|████████▌ | 12/14 [00:56<00:15, 7.74s/it]21310/34019 docs
Fetching 14 files: 93%|█████████▎| 13/14 [01:00<00:07, 7.02s/it]Fetching 14 files: 100%|██████████| 14/14 [01:02<00:00, 4.46s/it]
Fetching 18 files: 0%| | 0/18 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 18 files: 6%|▌ | 1/18 [00:00<00:02, 8.41it/s]Fetching 18 files: 28%|██▊ | 5/18 [00:00<00:01, 12.06it/s]22910/34019 docs
24510/34019 docs
26110/34019 docs
Fetching 18 files: 33%|███▎ | 6/18 [00:20<00:00, 12.06it/s]27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 22:15:48.788527690 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:15:48 patch 0.6B
Fetching 18 files: 39%|███▉ | 7/18 [00:47<01:38, 8.98s/it]Fetching 18 files: 44%|████▍ | 8/18 [00:47<01:11, 7.11s/it]Fetching 18 files: 50%|█████ | 9/18 [00:48<00:50, 5.60s/it]Fetching 18 files: 61%|██████ | 11/18 [00:48<00:24, 3.46s/it]Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 24%|██▎ | 73/310 [00:00<00:00, 729.67it/s]Loading weights: 47%|████▋ | 146/310 [00:00<00:00, 466.36it/s]Loading weights: 64%|██████▍ | 198/310 [00:00<00:00, 470.29it/s]Loading weights: 80%|████████ | 249/310 [00:00<00:00, 367.45it/s]Loading weights: 94%|█████████▎| 290/310 [00:00<00:00, 296.28it/s]Loading weights: 100%|██████████| 310/310 [00:00<00:00, 346.15it/s]
Fetching 18 files: 67%|██████▋ | 12/18 [00:59<00:29, 4.94s/it]Fetching 18 files: 72%|███████▏ | 13/18 [01:09<00:31, 6.36s/it]Fetching 18 files: 78%|███████▊ | 14/18 [01:16<00:25, 6.29s/it]Fetching 18 files: 83%|████████▎ | 15/18 [01:19<00:16, 5.65s/it]100/3000 samples
Fetching 18 files: 89%|████████▉ | 16/18 [01:29<00:13, 6.66s/it]Fetching 18 files: 94%|█████████▍| 17/18 [01:36<00:06, 6.80s/it]200/3000 samples
Fetching 18 files: 100%|██████████| 18/18 [01:41<00:00, 6.31s/it]Fetching 18 files: 100%|██████████| 18/18 [01:41<00:00, 5.64s/it]
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1014 22:24:11.909273151 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: setting up run m39t7lyy
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261014_222414-m39t7lyy
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-0.6B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: 🚀 View run at https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/m39t7lyy
wandb: updating run metadata; uploading data
wandb: uploading wandb-summary.json; uploading config.yaml; uploading wandb-metadata.json; uploading requirements.txt
wandb:
wandb: Run history:
wandb: global/disruption ▂▄▅▂▂▁▂▂▂▁▂▁▂▁▂▂▂▄▄▅▅▆▅▆▇▇▇█▆
wandb: global/dloss ▁▁▂▁▁▁▁▁▂▁▁▁▂▁▂▂▁▂▂▃▃▄▄▅▅▆▆▆█
wandb: global/kl ▁▁▂▁▁▁▁▁▂▁▁▂▂▁▂▂▁▂▂▃▃▄▄▅▅▆▆▆█
wandb: global/retained █▇▆▄▄▄▄▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁
wandb: mean/disruption ▁▁▁▁▂▂▂▂▂▂▂▂▂▂▃▃▃▄▅▆▆▇▇▇▇███
wandb: mean/dloss ▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▃▃▄▄▅▅▆▆▆▇█
wandb: mean/kl ▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▃▃▄▄▅▅▆▆▆▇█
wandb: mean/retained ▁▁▂▂▃▄▄▄▄▅▄▄▄▄▅▆▇▇██▇▇▇▆▆▆▆▆
wandb: patch_site ▁▁▁▂▂▂▃▃▃▃▃▄▄▄▅▅▅▅▅▆▆▆▇▇▇▇▇██
wandb: relative_depth ▁▁▂▂▂▂▃▃▃▃▄▄▄▄▅▅▅▅▅▆▆▆▆▇▇▇▇██
wandb: +5 ...
wandb:
wandb: Run summary:
wandb: global/disruption 1.34938
wandb: global/dloss 6.64852
wandb: global/kl 6.55138
wandb: global/retained 1
wandb: mean/disruption 1
wandb: mean/disruption_reaches_half_at_depth 0.64286
wandb: mean/dloss 4.30993
wandb: mean/kl 4.1832
wandb: mean/retained 1
wandb: patch_site 28
wandb: +6 ...
wandb:
wandb: 🚀 View run Qwen3-0.6B-Base at: https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/m39t7lyy
wandb: ⭐️ View project at: https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: Synced 4 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20261014_222414-m39t7lyy/logs
== 22:24:16 means 1.7B
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 35%|███▍ | 107/310 [00:00<00:00, 1020.69it/s]Loading weights: 68%|██████▊ | 210/310 [00:00<00:00, 705.53it/s] Loading weights: 92%|█████████▏| 286/310 [00:00<00:00, 695.21it/s]Loading weights: 100%|██████████| 310/310 [00:00<00:00, 734.71it/s]
34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
6591/34019 docs
10533/34019 docs
13510/34019 docs
15912/34019 docs
17940/34019 docs
19703/34019 docs
21310/34019 docs
22910/34019 docs
24510/34019 docs
26110/34019 docs
27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 22:27:35.036412337 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:27:36 patch 1.7B
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 0%| | 1/310 [00:00<01:03, 4.85it/s]Loading weights: 12%|█▏ | 36/310 [00:00<00:02, 109.01it/s]Loading weights: 18%|█▊ | 56/310 [00:00<00:01, 136.37it/s]Loading weights: 23%|██▎ | 72/310 [00:00<00:01, 142.79it/s]Loading weights: 28%|██▊ | 88/310 [00:00<00:01, 147.30it/s]Loading weights: 34%|███▍ | 106/310 [00:00<00:01, 157.07it/s]Loading weights: 40%|████ | 124/310 [00:00<00:01, 161.37it/s]Loading weights: 46%|████▌ | 143/310 [00:01<00:01, 166.59it/s]Loading weights: 53%|█████▎ | 163/310 [00:01<00:01, 141.04it/s]Loading weights: 58%|█████▊ | 181/310 [00:01<00:00, 149.90it/s]Loading weights: 64%|██████▎ | 197/310 [00:01<00:00, 151.15it/s]Loading weights: 69%|██████▊ | 213/310 [00:01<00:00, 122.26it/s]Loading weights: 73%|███████▎ | 227/310 [00:01<00:00, 124.85it/s]Loading weights: 78%|███████▊ | 242/310 [00:01<00:00, 131.06it/s]Loading weights: 83%|████████▎ | 256/310 [00:01<00:00, 107.33it/s]Loading weights: 86%|████████▋ | 268/310 [00:02<00:00, 87.74it/s] Loading weights: 90%|█████████ | 279/310 [00:02<00:00, 76.24it/s]Loading weights: 93%|█████████▎| 288/310 [00:02<00:00, 78.53it/s]Loading weights: 96%|█████████▌| 297/310 [00:02<00:00, 57.14it/s]Loading weights: 98%|█████████▊| 305/310 [00:02<00:00, 60.82it/s]Loading weights: 100%|██████████| 310/310 [00:02<00:00, 106.88it/s]
100/3000 samples
200/3000 samples
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1014 22:37:36.191479552 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261014_223740-l5qxnakn
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-1.7B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: 🚀 View run at https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/l5qxnakn
wandb: updating run metadata; uploading data
wandb: uploading requirements.txt; uploading wandb-summary.json; uploading wandb-metadata.json
wandb:
wandb: Run history:
wandb: global/disruption ▂▄▄▂▁▁▁▁▁▁▂▂▂▂▂▁▁▄▆▆▇███████▄
wandb: global/dloss ▁▂▂▁▁▁▁▁▁▂▁▂▂▂▁▂▂▂▂▃▄▅▅▆▇▇▇▆█
wandb: global/kl ▁▂▂▁▁▁▁▁▁▂▁▂▂▂▁▂▂▂▂▃▄▅▅▆▇▇▇▆█
wandb: global/retained █▆▅▄▄▄▃▃▃▃▂▂▂▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁
wandb: mean/disruption ▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▃▃▄▅▆▇▇▇████▇
wandb: mean/dloss ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▃▃▄▅▅▅▆▆▇▇█
wandb: mean/kl ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▄▅▅▅▆▆▇▇█
wandb: mean/retained ▁▁▃▃▅▅▆▆▆▇▆▅▅▅▆▇▇▇█▇▆▆▅▄▄▄▄▅
wandb: patch_site ▁▁▁▂▂▂▃▃▃▃▃▄▄▄▅▅▅▅▅▆▆▆▇▇▇▇▇██
wandb: relative_depth ▁▁▂▂▂▂▃▃▃▃▄▄▄▄▅▅▅▅▅▆▆▆▆▇▇▇▇██
wandb: +5 ...
wandb:
wandb: Run summary:
wandb: global/disruption 1.32251
wandb: global/dloss 6.75708
wandb: global/kl 6.73423
wandb: global/retained 1
wandb: mean/disruption 1
wandb: mean/disruption_reaches_half_at_depth 0.64286
wandb: mean/dloss 4.53292
wandb: mean/kl 4.48372
wandb: mean/retained 1
wandb: patch_site 28
wandb: +6 ...
wandb:
wandb: 🚀 View run Qwen3-1.7B-Base at: https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/l5qxnakn
wandb: ⭐️ View project at: https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: Synced 4 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20261014_223740-l5qxnakn/logs
== 22:37:43 means 4B
Loading weights: 0%| | 0/398 [00:00<?, ?it/s]Loading weights: 18%|█▊ | 73/398 [00:00<00:00, 727.67it/s]Loading weights: 37%|███▋ | 146/398 [00:00<00:00, 399.18it/s]Loading weights: 49%|████▊ | 194/398 [00:00<00:00, 286.67it/s]Loading weights: 100%|██████████| 398/398 [00:00<00:00, 650.29it/s]
34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
6591/34019 docs
10533/34019 docs
13510/34019 docs
15912/34019 docs
17940/34019 docs
19703/34019 docs
21310/34019 docs
22910/34019 docs
24510/34019 docs
26110/34019 docs
27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 22:43:45.605485872 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:43:45 patch 4B
Loading weights: 0%| | 0/398 [00:00<?, ?it/s]Loading weights: 0%| | 1/398 [00:00<00:43, 9.15it/s]Loading weights: 5%|▌ | 21/398 [00:00<00:04, 76.16it/s]Loading weights: 9%|▉ | 36/398 [00:00<00:03, 98.53it/s]Loading weights: 12%|█▏ | 47/398 [00:00<00:04, 79.62it/s]Loading weights: 14%|█▍ | 56/398 [00:00<00:05, 64.44it/s]Loading weights: 17%|█▋ | 66/398 [00:00<00:05, 59.86it/s]Loading weights: 18%|█▊ | 73/398 [00:01<00:05, 60.56it/s]Loading weights: 20%|██ | 80/398 [00:01<00:05, 62.29it/s]Loading weights: 22%|██▏ | 88/398 [00:01<00:06, 45.50it/s]Loading weights: 25%|██▍ | 98/398 [00:01<00:05, 55.77it/s]Loading weights: 26%|██▋ | 105/398 [00:01<00:06, 47.83it/s]Loading weights: 28%|██▊ | 113/398 [00:02<00:06, 44.49it/s]Loading weights: 30%|███ | 121/398 [00:02<00:06, 44.56it/s]Loading weights: 32%|███▏ | 126/398 [00:02<00:06, 44.60it/s]Loading weights: 33%|███▎ | 131/398 [00:02<00:08, 32.21it/s]Loading weights: 35%|███▌ | 141/398 [00:02<00:06, 38.11it/s]Loading weights: 37%|███▋ | 146/398 [00:02<00:06, 39.79it/s]Loading weights: 39%|███▊ | 154/398 [00:03<00:07, 34.47it/s]Loading weights: 41%|████ | 163/398 [00:03<00:05, 43.89it/s]Loading weights: 42%|████▏ | 169/398 [00:03<00:05, 39.51it/s]Loading weights: 45%|████▍ | 178/398 [00:03<00:04, 48.38it/s]Loading weights: 46%|████▌ | 184/398 [00:03<00:05, 41.62it/s]Loading weights: 48%|████▊ | 190/398 [00:03<00:04, 45.09it/s]Loading weights: 49%|████▉ | 196/398 [00:04<00:04, 40.76it/s]Loading weights: 51%|█████ | 201/398 [00:04<00:04, 42.19it/s]Loading weights: 53%|█████▎ | 209/398 [00:04<00:05, 35.12it/s]Loading weights: 54%|█████▍ | 214/398 [00:04<00:04, 37.79it/s]Loading weights: 56%|█████▌ | 222/398 [00:04<00:04, 38.76it/s]Loading weights: 57%|█████▋ | 227/398 [00:05<00:05, 29.18it/s]Loading weights: 58%|█████▊ | 231/398 [00:05<00:06, 26.39it/s]Loading weights: 59%|█████▉ | 236/398 [00:05<00:06, 26.26it/s]Loading weights: 62%|██████▏ | 245/398 [00:05<00:04, 30.94it/s]Loading weights: 64%|██████▎ | 253/398 [00:06<00:04, 29.99it/s]Loading weights: 65%|██████▍ | 257/398 [00:06<00:04, 31.48it/s]Loading weights: 66%|██████▋ | 264/398 [00:06<00:04, 32.42it/s]Loading weights: 69%|██████▉ | 276/398 [00:06<00:02, 47.65it/s]Loading weights: 71%|███████ | 282/398 [00:06<00:03, 35.61it/s]Loading weights: 72%|███████▏ | 287/398 [00:07<00:03, 28.86it/s]Loading weights: 73%|███████▎ | 291/398 [00:07<00:03, 30.30it/s]Loading weights: 75%|███████▍ | 297/398 [00:07<00:02, 35.62it/s]Loading weights: 77%|███████▋ | 308/398 [00:07<00:02, 36.12it/s]Loading weights: 80%|███████▉ | 318/398 [00:07<00:01, 47.05it/s]Loading weights: 81%|████████▏ | 324/398 [00:07<00:01, 41.13it/s]Loading weights: 84%|████████▎ | 333/398 [00:07<00:01, 49.67it/s]Loading weights: 86%|████████▋ | 344/398 [00:08<00:01, 45.10it/s]Loading weights: 88%|████████▊ | 350/398 [00:08<00:01, 46.96it/s]Loading weights: 89%|████████▉ | 356/398 [00:08<00:01, 41.61it/s]Loading weights: 91%|█████████ | 363/398 [00:08<00:00, 39.13it/s]Loading weights: 93%|█████████▎| 372/398 [00:08<00:00, 47.97it/s]Loading weights: 95%|█████████▍| 378/398 [00:09<00:00, 31.57it/s]Loading weights: 97%|█████████▋| 385/398 [00:09<00:00, 32.68it/s]Loading weights: 100%|██████████| 398/398 [00:09<00:00, 42.27it/s]
100/3000 samples
200/3000 samples
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1014 23:03:26.845779633 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: setting up run 9g4x3o1r
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261014_230329-9g4x3o1r
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-4B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: 🚀 View run at https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/9g4x3o1r
wandb: updating run metadata; uploading data
wandb: uploading requirements.txt; uploading wandb-summary.json
wandb:
wandb: Run history:
wandb: global/disruption ▃▄▃▂▁▁▂▁▁▁▂▁▁▂▂▁▃▁▃▂▂▃▄▅▆▆▅▆▇▇███▆▅▆▄
wandb: global/dloss ▁▂▂▂▁▁▁▁▁▁▂▁▁▂▂▂▂▂▂▂▂▂▂▂▂▃▃▄▄▅▆▆▇▇▇▆█
wandb: global/kl ▁▂▂▂▁▁▁▁▁▁▂▁▁▂▂▂▂▂▂▂▂▂▂▂▃▃▃▄▄▅▆▆▇▇▇▆█
wandb: global/retained █▄▃▃▂▂▂▁▁▁▂▂▂▂▂▂▁▂▂▁▂▂▂▂▂▂▂▂▂▁▁▁▁▁▁▁
wandb: mean/disruption ▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▄▅▅▆▆▇▇████▇█▇
wandb: mean/dloss ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▄▄▄▅▅▆▆▆▇█
wandb: mean/kl ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▄▄▄▅▅▆▆▆▇█
wandb: mean/retained ▂▁▁▁▁▂▂▂▂▂▂▂▃▃▃▃▂▃▃▃▄▅▆▇█████▇▇▇▇▆▇▇
wandb: patch_site ▁▁▁▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇███
wandb: relative_depth ▁▁▁▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇███
wandb: +5 ...
wandb:
wandb: Run summary:
wandb: global/disruption 1.37115
wandb: global/dloss 7.38363
wandb: global/kl 7.3311
wandb: global/retained 1
wandb: mean/disruption 1
wandb: mean/disruption_reaches_half_at_depth 0.63889
wandb: mean/dloss 4.96598
wandb: mean/kl 4.90221
wandb: mean/retained 1
wandb: patch_site 36
wandb: +6 ...
wandb:
wandb: 🚀 View run Qwen3-4B-Base at: https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/9g4x3o1r
wandb: ⭐️ View project at: https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: Synced 4 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20261014_230329-9g4x3o1r/logs
== 23:03:31 means 8B
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 0%| | 0/399 [00:00<?, ?it/s]Loading weights: 27%|██▋ | 106/399 [00:00<00:00, 1011.46it/s]Loading weights: 52%|█████▏ | 208/399 [00:00<00:00, 671.40it/s] Loading weights: 71%|███████ | 282/399 [00:00<00:00, 509.62it/s]Loading weights: 100%|██████████| 399/399 [00:00<00:00, 750.00it/s]
34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
6591/34019 docs
10533/34019 docs
13510/34019 docs
15912/34019 docs
17940/34019 docs
19703/34019 docs
21310/34019 docs
22910/34019 docs
24510/34019 docs
26110/34019 docs
27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 23:12:17.196354440 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 23:12:18 patch 8B
Loading weights: 0%| | 0/399 [00:00<?, ?it/s]Loading weights: 0%| | 1/399 [00:01<07:58, 1.20s/it]Loading weights: 1%| | 2/399 [00:01<04:03, 1.63it/s]Loading weights: 2%|▏ | 6/399 [00:01<01:02, 6.25it/s]Loading weights: 4%|▍ | 15/399 [00:01<00:20, 18.48it/s]Loading weights: 7%|▋ | 26/399 [00:01<00:14, 25.54it/s]Loading weights: 9%|▊ | 34/399 [00:02<00:12, 29.56it/s]Loading weights: 10%|▉ | 39/399 [00:02<00:11, 32.50it/s]Loading weights: 11%|█▏ | 45/399 [00:02<00:12, 27.78it/s]Loading weights: 13%|█▎ | 50/399 [00:02<00:11, 31.32it/s]Loading weights: 14%|█▍ | 56/399 [00:02<00:11, 30.83it/s]Loading weights: 15%|█▌ | 60/399 [00:02<00:10, 32.06it/s]Loading weights: 18%|█▊ | 70/399 [00:03<00:11, 28.87it/s]Loading weights: 20%|██ | 81/399 [00:03<00:10, 31.21it/s]Loading weights: 23%|██▎ | 92/399 [00:04<00:10, 29.73it/s]Loading weights: 25%|██▌ | 100/399 [00:04<00:10, 29.08it/s]Loading weights: 26%|██▌ | 104/399 [00:04<00:10, 26.86it/s]Loading weights: 29%|██▊ | 114/399 [00:04<00:08, 32.52it/s]Loading weights: 30%|███ | 120/399 [00:04<00:08, 32.16it/s]Loading weights: 31%|███ | 124/399 [00:05<00:08, 32.94it/s]Loading weights: 32%|███▏ | 128/399 [00:05<00:08, 33.76it/s]Loading weights: 33%|███▎ | 133/399 [00:05<00:09, 26.99it/s]Loading weights: 34%|███▍ | 136/399 [00:05<00:11, 23.43it/s]Loading weights: 36%|███▌ | 142/399 [00:05<00:08, 29.61it/s]Loading weights: 37%|███▋ | 146/399 [00:05<00:07, 31.67it/s]Loading weights: 38%|███▊ | 150/399 [00:05<00:07, 33.02it/s]Loading weights: 40%|███▉ | 158/399 [00:06<00:09, 26.06it/s]Loading weights: 42%|████▏ | 169/399 [00:06<00:08, 26.66it/s]Loading weights: 44%|████▍ | 175/399 [00:07<00:10, 22.16it/s]Loading weights: 47%|████▋ | 186/399 [00:07<00:06, 32.68it/s]Loading weights: 48%|████▊ | 191/399 [00:07<00:06, 30.52it/s]Loading weights: 49%|████▉ | 197/399 [00:07<00:06, 30.56it/s]Loading weights: 51%|█████ | 202/399 [00:07<00:06, 28.90it/s]Loading weights: 53%|█████▎ | 210/399 [00:08<00:07, 25.03it/s]Loading weights: 54%|█████▎ | 214/399 [00:08<00:07, 24.01it/s]Loading weights: 55%|█████▍ | 219/399 [00:08<00:08, 21.25it/s]Loading weights: 58%|█████▊ | 232/399 [00:08<00:04, 35.65it/s]Loading weights: 60%|█████▉ | 238/399 [00:09<00:05, 27.25it/s]Loading weights: 61%|██████ | 243/399 [00:09<00:05, 26.76it/s]Loading weights: 63%|██████▎ | 252/399 [00:09<00:04, 31.44it/s]Loading weights: 64%|██████▍ | 256/399 [00:09<00:04, 32.70it/s]Loading weights: 65%|██████▌ | 260/399 [00:10<00:06, 21.86it/s]Loading weights: 69%|██████▊ | 274/399 [00:10<00:03, 34.15it/s]Loading weights: 70%|██████▉ | 279/399 [00:10<00:03, 31.97it/s]Loading weights: 71%|███████ | 283/399 [00:10<00:03, 32.93it/s]Loading weights: 73%|███████▎ | 290/399 [00:11<00:05, 20.76it/s]Loading weights: 75%|███████▍ | 298/399 [00:11<00:04, 24.74it/s]Loading weights: 76%|███████▌ | 302/399 [00:11<00:04, 23.88it/s]Loading weights: 77%|███████▋ | 309/399 [00:11<00:02, 30.24it/s]Loading weights: 78%|███████▊ | 313/399 [00:11<00:03, 27.27it/s]Loading weights: 81%|████████ | 322/399 [00:12<00:02, 32.52it/s]Loading weights: 82%|████████▏ | 326/399 [00:12<00:03, 22.22it/s]Loading weights: 82%|████████▏ | 329/399 [00:12<00:03, 20.65it/s]Loading weights: 83%|████████▎ | 332/399 [00:12<00:03, 21.87it/s]Loading weights: 84%|████████▍ | 336/399 [00:12<00:02, 25.03it/s]Loading weights: 86%|████████▋ | 345/399 [00:13<00:02, 23.81it/s]Loading weights: 88%|████████▊ | 353/399 [00:13<00:01, 28.25it/s]Loading weights: 89%|████████▉ | 357/399 [00:13<00:01, 22.58it/s]Loading weights: 90%|█████████ | 360/399 [00:13<00:01, 23.65it/s]Loading weights: 91%|█████████ | 364/399 [00:14<00:02, 17.43it/s]Loading weights: 92%|█████████▏| 367/399 [00:14<00:01, 19.09it/s]Loading weights: 93%|█████████▎| 373/399 [00:14<00:01, 22.54it/s]Loading weights: 94%|█████████▍| 376/399 [00:14<00:00, 23.78it/s]Loading weights: 95%|█████████▍| 379/399 [00:14<00:00, 24.77it/s]Loading weights: 96%|█████████▌| 382/399 [00:14<00:00, 25.60it/s]Loading weights: 97%|█████████▋| 387/399 [00:15<00:00, 26.09it/s]Loading weights: 98%|█████████▊| 390/399 [00:15<00:00, 16.25it/s]Loading weights: 100%|██████████| 399/399 [00:15<00:00, 25.74it/s]
100/3000 samples
200/3000 samples
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1014 23:41:31.290283078 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: setting up run 53pmlbk2
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261014_234133-53pmlbk2
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-8B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: 🚀 View run at https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/53pmlbk2
wandb: updating run metadata; uploading data
wandb: uploading requirements.txt; uploading wandb-summary.json; uploading config.yaml; uploading wandb-metadata.json
wandb:
wandb: Run history:
wandb: global/disruption ▃▂▂▂▁▁▁▁▁▁▁▁▁▁▂▁▂▂▂▂▃▄▄▅▆▆▆▇███▇█▆▅▄▂
wandb: global/dloss ▂▂▁▂▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▄▄▅▆▆▇▇█▇▆█
wandb: global/kl ▁▂▁▂▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▂▃▃▄▄▅▆▆▇▇█▇▆█
wandb: global/retained █▆▅▅▄▄▃▃▃▃▃▃▃▃▃▂▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂
wandb: mean/disruption ▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▂▂▃▃▃▄▅▆▆▇▇█████▇▇▆
wandb: mean/dloss ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▄▄▅▅▅▆▆▆▇█
wandb: mean/kl ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▄▄▅▅▅▆▆▇▇█
wandb: mean/retained ▃▃▃▃▃▄▄▄▄▄▄▅▅▅▆▆▁▁▁▁▂▂▃▄▅▆▇▇▇▇▇▇▇▇██
wandb: patch_site ▁▁▁▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇███
wandb: relative_depth ▁▁▁▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇███
wandb: +5 ...
wandb:
wandb: Run summary:
wandb: global/disruption 1.37855
wandb: global/dloss 7.32931
wandb: global/kl 7.26291
wandb: global/retained 1.0
wandb: mean/disruption 1
wandb: mean/disruption_reaches_half_at_depth 0.61111
wandb: mean/dloss 4.97322
wandb: mean/kl 4.88872
wandb: mean/retained 1
wandb: patch_site 36
wandb: +6 ...
wandb:
wandb: 🚀 View run Qwen3-8B-Base at: https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/53pmlbk2
wandb: ⭐️ View project at: https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: Synced 4 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20261014_234133-53pmlbk2/logs
== 23:41:35 means 14B
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 0%| | 0/443 [00:00<?, ?it/s]Loading weights: 20%|█▉ | 88/443 [00:00<00:00, 845.23it/s]Loading weights: 39%|███▉ | 173/443 [00:00<00:00, 578.11it/s]Loading weights: 53%|█████▎ | 236/443 [00:00<00:00, 525.19it/s]Loading weights: 66%|██████▌ | 291/443 [00:00<00:00, 490.55it/s]Loading weights: 77%|███████▋ | 342/443 [00:00<00:00, 469.61it/s]Loading weights: 89%|████████▉ | 394/443 [00:00<00:00, 483.55it/s]Loading weights: 100%|██████████| 443/443 [00:00<00:00, 531.38it/s]
34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
6591/34019 docs
10533/34019 docs
13510/34019 docs
15912/34019 docs
17940/34019 docs
19703/34019 docs
21310/34019 docs
22910/34019 docs
24510/34019 docs
26110/34019 docs
27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 23:55:40.058281140 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 23:55:41 patch 14B
Loading weights: 0%| | 0/443 [00:00<?, ?it/s]Loading weights: 0%| | 1/443 [00:04<36:39, 4.98s/it]Loading weights: 6%|▋ | 28/443 [00:05<00:58, 7.12it/s]Loading weights: 7%|▋ | 32/443 [00:05<00:50, 8.11it/s]Loading weights: 8%|▊ | 37/443 [00:06<00:52, 7.72it/s]Loading weights: 10%|█ | 45/443 [00:06<00:35, 11.20it/s]Loading weights: 11%|█▏ | 50/443 [00:07<00:46, 8.42it/s]Loading weights: 13%|█▎ | 56/443 [00:07<00:35, 10.81it/s]Loading weights: 14%|█▎ | 60/443 [00:08<00:43, 8.77it/s]Loading weights: 15%|█▌ | 67/443 [00:08<00:31, 12.06it/s]Loading weights: 16%|█▌ | 70/443 [00:09<00:44, 8.45it/s]Loading weights: 16%|█▋ | 73/443 [00:09<00:38, 9.70it/s]Loading weights: 18%|█▊ | 81/443 [00:10<00:38, 9.40it/s]Loading weights: 19%|█▊ | 83/443 [00:10<00:36, 9.76it/s]Loading weights: 21%|██ | 92/443 [00:11<00:34, 10.27it/s]Loading weights: 21%|██ | 94/443 [00:11<00:32, 10.78it/s]Loading weights: 23%|██▎ | 100/443 [00:11<00:22, 15.10it/s]Loading weights: 23%|██▎ | 103/443 [00:12<00:31, 10.75it/s]Loading weights: 24%|██▍ | 106/443 [00:12<00:29, 11.45it/s]Loading weights: 26%|██▌ | 114/443 [00:13<00:27, 11.77it/s]Loading weights: 26%|██▌ | 116/443 [00:13<00:36, 8.89it/s]Loading weights: 28%|██▊ | 125/443 [00:14<00:33, 9.40it/s]Loading weights: 29%|██▊ | 127/443 [00:14<00:32, 9.81it/s]Loading weights: 31%|███ | 136/443 [00:15<00:29, 10.47it/s]Loading weights: 31%|███ | 138/443 [00:15<00:30, 10.03it/s]Loading weights: 33%|███▎ | 145/443 [00:15<00:21, 13.92it/s]Loading weights: 33%|███▎ | 147/443 [00:16<00:30, 9.69it/s]Loading weights: 35%|███▍ | 155/443 [00:16<00:20, 14.21it/s]Loading weights: 36%|███▌ | 158/443 [00:17<00:27, 10.50it/s]Loading weights: 36%|███▌ | 160/443 [00:17<00:35, 7.90it/s]Loading weights: 38%|███▊ | 169/443 [00:18<00:21, 12.96it/s]Loading weights: 40%|███▉ | 175/443 [00:18<00:17, 15.52it/s]Loading weights: 40%|████ | 178/443 [00:18<00:16, 15.98it/s]Loading weights: 41%|████ | 181/443 [00:19<00:26, 9.80it/s]Loading weights: 43%|████▎ | 191/443 [00:20<00:23, 10.66it/s]Loading weights: 44%|████▎ | 193/443 [00:20<00:23, 10.83it/s]Loading weights: 45%|████▍ | 199/443 [00:20<00:16, 15.00it/s]Loading weights: 46%|████▌ | 202/443 [00:21<00:25, 9.45it/s]Loading weights: 46%|████▌ | 204/443 [00:21<00:23, 10.17it/s]Loading weights: 48%|████▊ | 213/443 [00:22<00:23, 9.73it/s]Loading weights: 51%|█████ | 224/443 [00:23<00:21, 10.10it/s]Loading weights: 53%|█████▎ | 235/443 [00:24<00:17, 11.78it/s]Loading weights: 53%|█████▎ | 237/443 [00:24<00:17, 11.85it/s]Loading weights: 56%|█████▌ | 246/443 [00:25<00:16, 12.01it/s]Loading weights: 56%|█████▌ | 248/443 [00:25<00:17, 11.13it/s]Loading weights: 58%|█████▊ | 257/443 [00:25<00:15, 12.02it/s]Loading weights: 58%|█████▊ | 259/443 [00:26<00:16, 11.18it/s]Loading weights: 60%|██████ | 268/443 [00:27<00:16, 10.65it/s]Loading weights: 61%|██████ | 270/443 [00:27<00:17, 10.15it/s]Loading weights: 63%|██████▎ | 279/443 [00:28<00:15, 10.29it/s]Loading weights: 63%|██████▎ | 281/443 [00:28<00:17, 9.52it/s]Loading weights: 65%|██████▌ | 290/443 [00:29<00:14, 10.79it/s]Loading weights: 66%|██████▌ | 292/443 [00:29<00:13, 11.01it/s]Loading weights: 67%|██████▋ | 298/443 [00:29<00:09, 15.29it/s]Loading weights: 68%|██████▊ | 301/443 [00:30<00:14, 10.12it/s]Loading weights: 68%|██████▊ | 303/443 [00:30<00:13, 10.42it/s]Loading weights: 70%|██████▉ | 309/443 [00:30<00:09, 13.75it/s]Loading weights: 70%|███████ | 312/443 [00:31<00:12, 10.37it/s]Loading weights: 71%|███████ | 314/443 [00:31<00:11, 11.09it/s]Loading weights: 72%|███████▏ | 320/443 [00:31<00:08, 15.26it/s]Loading weights: 73%|███████▎ | 323/443 [00:32<00:14, 8.40it/s]Loading weights: 73%|███████▎ | 325/443 [00:32<00:13, 9.02it/s]Loading weights: 75%|███████▌ | 334/443 [00:33<00:10, 10.17it/s]Loading weights: 77%|███████▋ | 342/443 [00:33<00:08, 11.73it/s]Loading weights: 78%|███████▊ | 345/443 [00:34<00:10, 9.57it/s]Loading weights: 80%|███████▉ | 353/443 [00:34<00:06, 12.91it/s]Loading weights: 80%|████████ | 356/443 [00:35<00:09, 9.11it/s]Loading weights: 81%|████████ | 358/443 [00:35<00:10, 8.42it/s]Loading weights: 83%|████████▎ | 367/443 [00:36<00:06, 12.48it/s]Loading weights: 83%|████████▎ | 369/443 [00:36<00:05, 13.13it/s]Loading weights: 85%|████████▍ | 375/443 [00:36<00:04, 14.25it/s]Loading weights: 85%|████████▌ | 378/443 [00:37<00:05, 10.86it/s]Loading weights: 87%|████████▋ | 386/443 [00:37<00:04, 13.72it/s]Loading weights: 88%|████████▊ | 389/443 [00:38<00:05, 9.14it/s]Loading weights: 88%|████████▊ | 391/443 [00:38<00:05, 9.80it/s]Loading weights: 90%|█████████ | 400/443 [00:39<00:03, 11.05it/s]Loading weights: 91%|█████████ | 402/443 [00:39<00:03, 10.79it/s]Loading weights: 91%|█████████ | 404/443 [00:39<00:03, 11.61it/s]Loading weights: 93%|█████████▎| 411/443 [00:40<00:03, 9.87it/s]Loading weights: 93%|█████████▎| 413/443 [00:40<00:02, 10.24it/s]Loading weights: 95%|█████████▍| 419/443 [00:40<00:01, 14.33it/s]Loading weights: 95%|█████████▌| 422/443 [00:41<00:02, 9.35it/s]Loading weights: 96%|█████████▌| 424/443 [00:41<00:02, 9.09it/s]Loading weights: 98%|█████████▊| 433/443 [00:42<00:00, 12.66it/s]Loading weights: 98%|█████████▊| 435/443 [00:42<00:00, 13.15it/s]Loading weights: 100%|██████████| 443/443 [00:42<00:00, 10.48it/s]
100/3000 samples
200/3000 samples
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1015 00:47:12.044365755 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: setting up run u8994v64
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261015_004715-u8994v64
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-14B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: 🚀 View run at https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/u8994v64
wandb: updating run metadata; uploading data
wandb: uploading wandb-metadata.json; uploading requirements.txt; uploading wandb-summary.json; uploading config.yaml
wandb:
wandb: Run history:
wandb: global/disruption ▂▂▂▁▁▂▁▁▁▁▁▁▁▂▁▁▁▁▁▂▁▁▂▂▂▂▃▄▅▆▇▅▄▃▃▂█▇▃▂
wandb: global/dloss ▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▁▁▁▁▁▁▁▁▁▁▁▁▂▂▃▃▄▄▄▅▅▅▅▅█
wandb: global/kl ▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▁▁▁▁▁▁▁▁▁▁▁▁▂▂▃▃▄▄▄▅▅▅▅▅█
wandb: global/retained █▆▅▄▄▃▃▃▂▂▂▂▂▂▂▂▂▂▂▁▁▁▁▁▂▂▂▂▁▁▁▁▁▁▁▁▁▁▁
wandb: mean/disruption ▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▂▃▃▃▄▅▅▆▇▇▇▆▆██▆▅
wandb: mean/dloss ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▃▃▃▄▄▄▅▅▅▆█
wandb: mean/kl ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▂▂▂▂▃▃▄▄▄▄▅▅▅▆█
wandb: mean/retained ▁▁▁▁▁▁▂▂▂▂▂▂▂▂▂▃▃▃▄▄▄▅▅▆▆▇█████▇▇▆▆▆▆▆▆
wandb: patch_site ▁▁▁▂▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▆▇▇▇▇▇▇██
wandb: relative_depth ▁▁▁▂▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▆▇▇▇▇▇▇██
wandb: +5 ...
wandb:
wandb: Run summary:
wandb: global/disruption 1.40922
wandb: global/dloss 9.30139
wandb: global/kl 9.25185
wandb: global/retained 1
wandb: mean/disruption 1
wandb: mean/disruption_reaches_half_at_depth 0.65
wandb: mean/dloss 6.10597
wandb: mean/kl 6.05207
wandb: mean/retained 1
wandb: patch_site 40
wandb: +6 ...
wandb:
wandb: 🚀 View run Qwen3-14B-Base at: https://wandb.ai/octahedral-systems/qwen3-self-correction/runs/u8994v64
wandb: ⭐️ View project at: https://wandb.ai/octahedral-systems/qwen3-self-correction
wandb: Synced 4 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s)
wandb: Find logs at: ./wandb/run-20261015_004715-u8994v64/logs
== 00:47:18 precision check 14B
Loading weights: 0%| | 0/443 [00:00<?, ?it/s]Loading weights: 16%|█▋ | 73/443 [00:00<00:00, 560.77it/s]Loading weights: 32%|███▏ | 143/443 [00:00<00:00, 586.02it/s]Loading weights: 46%|████▌ | 202/443 [00:00<00:00, 555.71it/s]Loading weights: 58%|█████▊ | 258/443 [00:00<00:00, 543.08it/s]Loading weights: 71%|███████ | 313/443 [00:00<00:00, 535.24it/s]Loading weights: 83%|████████▎ | 367/443 [00:00<00:00, 517.98it/s]Loading weights: 95%|█████████▍| 419/443 [00:00<00:00, 504.84it/s]Loading weights: 100%|██████████| 443/443 [00:00<00:00, 556.71it/s]
1042 docs, 300153 positions, 271 tokens with >= 100 occurrences (53% of positions)
[W1015 00:49:04.706419105 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
Loading weights: 0%| | 0/443 [00:00<?, ?it/s]Loading weights: 0%| | 1/443 [00:01<10:19, 1.40s/it]Loading weights: 0%| | 2/443 [00:01<04:40, 1.57it/s]Loading weights: 1%|▏ | 6/443 [00:01<01:12, 6.02it/s]Loading weights: 3%|▎ | 15/443 [00:01<00:30, 14.12it/s]Loading weights: 4%|▍ | 18/443 [00:02<00:44, 9.63it/s]Loading weights: 6%|▌ | 26/443 [00:02<00:27, 15.41it/s]Loading weights: 8%|▊ | 34/443 [00:03<00:24, 16.94it/s]Loading weights: 8%|▊ | 37/443 [00:03<00:24, 16.58it/s]Loading weights: 9%|▉ | 40/443 [00:03<00:22, 18.23it/s]Loading weights: 10%|█ | 45/443 [00:03<00:19, 20.05it/s]Loading weights: 11%|█ | 48/443 [00:03<00:21, 18.47it/s]Loading weights: 13%|█▎ | 56/443 [00:04<00:17, 21.61it/s]Loading weights: 13%|█▎ | 59/443 [00:04<00:21, 17.49it/s]Loading weights: 14%|█▍ | 61/443 [00:04<00:27, 14.10it/s]Loading weights: 14%|█▍ | 63/443 [00:04<00:28, 13.16it/s]Loading weights: 16%|█▌ | 70/443 [00:05<00:20, 18.65it/s]Loading weights: 18%|█▊ | 81/443 [00:05<00:19, 18.40it/s]Loading weights: 21%|██ | 92/443 [00:06<00:17, 19.78it/s]Loading weights: 23%|██▎ | 103/443 [00:06<00:17, 19.13it/s]Loading weights: 26%|██▌ | 113/443 [00:06<00:13, 24.33it/s]Loading weights: 26%|██▋ | 117/443 [00:07<00:17, 18.32it/s]Loading weights: 28%|██▊ | 122/443 [00:07<00:17, 17.89it/s]Loading weights: 28%|██▊ | 125/443 [00:08<00:18, 17.41it/s]Loading weights: 30%|███ | 133/443 [00:08<00:18, 16.75it/s]Loading weights: 31%|███ | 136/443 [00:08<00:17, 18.04it/s]Loading weights: 33%|███▎ | 144/443 [00:09<00:17, 17.34it/s]Loading weights: 33%|███▎ | 147/443 [00:09<00:17, 16.89it/s]Loading weights: 34%|███▎ | 149/443 [00:09<00:19, 15.33it/s]Loading weights: 34%|███▍ | 151/443 [00:09<00:23, 12.68it/s]Loading weights: 35%|███▍ | 153/443 [00:09<00:24, 12.05it/s]Loading weights: 36%|███▌ | 158/443 [00:10<00:16, 17.18it/s]Loading weights: 36%|███▋ | 161/443 [00:10<00:17, 16.58it/s]Loading weights: 38%|███▊ | 169/443 [00:10<00:16, 16.40it/s]Loading weights: 40%|███▉ | 177/443 [00:11<00:16, 16.23it/s]Loading weights: 41%|████ | 180/443 [00:11<00:14, 17.61it/s]Loading weights: 42%|████▏ | 188/443 [00:11<00:12, 20.66it/s]Loading weights: 43%|████▎ | 191/443 [00:12<00:14, 17.21it/s]Loading weights: 44%|████▎ | 193/443 [00:12<00:16, 15.62it/s]Loading weights: 45%|████▍ | 199/443 [00:12<00:14, 17.14it/s]Loading weights: 46%|████▌ | 202/443 [00:12<00:14, 16.60it/s]Loading weights: 47%|████▋ | 210/443 [00:12<00:10, 22.59it/s]Loading weights: 48%|████▊ | 213/443 [00:13<00:15, 14.63it/s]Loading weights: 49%|████▉ | 217/443 [00:13<00:17, 13.05it/s]Loading weights: 51%|█████ | 224/443 [00:14<00:12, 17.39it/s]Loading weights: 51%|█████ | 227/443 [00:14<00:12, 16.86it/s]Loading weights: 52%|█████▏ | 229/443 [00:14<00:12, 17.29it/s]Loading weights: 53%|█████▎ | 235/443 [00:14<00:10, 20.66it/s]Loading weights: 55%|█████▍ | 243/443 [00:14<00:08, 22.97it/s]Loading weights: 56%|█████▌ | 246/443 [00:15<00:10, 18.55it/s]Loading weights: 57%|█████▋ | 254/443 [00:15<00:09, 19.02it/s]Loading weights: 58%|█████▊ | 257/443 [00:15<00:10, 18.25it/s]Loading weights: 58%|█████▊ | 259/443 [00:15<00:09, 18.45it/s]Loading weights: 60%|██████ | 267/443 [00:15<00:06, 28.51it/s]Loading weights: 61%|██████ | 271/443 [00:16<00:08, 19.43it/s]Loading weights: 62%|██████▏ | 276/443 [00:16<00:07, 21.01it/s]Loading weights: 63%|██████▎ | 279/443 [00:16<00:10, 15.24it/s]Loading weights: 64%|██████▍ | 285/443 [00:17<00:09, 16.88it/s]Loading weights: 65%|██████▌ | 290/443 [00:17<00:10, 15.21it/s]Loading weights: 66%|██████▌ | 292/443 [00:17<00:09, 15.68it/s]Loading weights: 67%|██████▋ | 296/443 [00:18<00:13, 11.18it/s]Loading weights: 68%|██████▊ | 301/443 [00:18<00:10, 13.91it/s]Loading weights: 69%|██████▉ | 307/443 [00:18<00:07, 17.41it/s]Loading weights: 70%|██████▉ | 310/443 [00:18<00:07, 16.84it/s]Loading weights: 70%|███████ | 312/443 [00:18<00:07, 17.26it/s]Loading weights: 71%|███████ | 314/443 [00:19<00:07, 17.40it/s]Loading weights: 72%|███████▏ | 318/443 [00:19<00:12, 10.18it/s]Loading weights: 73%|███████▎ | 323/443 [00:19<00:08, 14.54it/s]Loading weights: 75%|███████▌ | 334/443 [00:20<00:07, 14.22it/s]Loading weights: 78%|███████▊ | 345/443 [00:21<00:07, 14.00it/s]Loading weights: 79%|███████▉ | 349/443 [00:21<00:06, 13.82it/s]Loading weights: 83%|████████▎ | 367/443 [00:22<00:04, 18.49it/s]Loading weights: 85%|████████▍ | 375/443 [00:23<00:03, 17.77it/s]Loading weights: 85%|████████▌ | 378/443 [00:23<00:03, 17.47it/s]Loading weights: 88%|████████▊ | 389/443 [00:23<00:02, 18.92it/s]Loading weights: 90%|█████████ | 400/443 [00:24<00:02, 17.59it/s]Loading weights: 92%|█████████▏| 407/443 [00:24<00:01, 21.53it/s]Loading weights: 93%|█████████▎| 411/443 [00:24<00:01, 18.18it/s]Loading weights: 95%|█████████▍| 420/443 [00:25<00:00, 24.92it/s]Loading weights: 96%|█████████▌| 425/443 [00:25<00:00, 18.79it/s]Loading weights: 97%|█████████▋| 431/443 [00:25<00:00, 21.05it/s]Loading weights: 98%|█████████▊| 435/443 [00:26<00:00, 15.61it/s]Loading weights: 100%|██████████| 443/443 [00:26<00:00, 16.85it/s]
1042 docs, 300153 positions, 271 tokens with >= 100 occurrences (53% of positions)
[W1015 00:53:23.505218158 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
[W1015 00:53:26.795152847 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 00:53:26 plot
== Qwen3-0.6B-Base (3000 samples, float32; values after the last block)
max |diff|^2 at sites before the patch site: 0 (must be 0)
mean-ablation disruption first reaches 0.5 at site 18 (relative depth 0.64)
-- mean ablation
site disruption (median) retained along KL dloss
0 nan 0.000 nan nan 0.000 0.000
1 0.030 0.002 0.30 0.00 0.034 0.031
2 0.050 0.004 0.32 0.00 0.066 0.064
3 0.071 0.010 0.43 0.00 0.099 0.093
4 0.089 0.016 0.46 0.00 0.128 0.121
5 0.109 0.027 0.63 0.01 0.163 0.153
6 0.134 0.044 0.65 0.01 0.213 0.206
7 0.148 0.053 0.70 0.01 0.229 0.226
8 0.164 0.068 0.68 0.01 0.254 0.250
9 0.175 0.078 0.72 0.01 0.264 0.262
10 0.194 0.098 0.79 0.01 0.293 0.294
11 0.212 0.119 0.76 0.02 0.315 0.318
12 0.219 0.129 0.74 0.02 0.327 0.330
13 0.226 0.141 0.75 0.02 0.340 0.343
14 0.235 0.147 0.77 0.02 0.359 0.365
15 0.264 0.179 0.89 0.03 0.414 0.422
16 0.348 0.261 1.02 0.04 0.578 0.588
17 0.394 0.308 1.07 0.05 0.713 0.722
18 0.510 0.429 1.15 0.08 1.065 1.084
19 0.645 0.586 1.25 0.11 1.464 1.508
20 0.729 0.681 1.23 0.16 1.773 1.832
21 0.803 0.763 1.18 0.22 2.075 2.144
22 0.843 0.812 1.11 0.29 2.246 2.323
23 0.910 0.883 1.06 0.36 2.519 2.612
24 0.946 0.919 0.99 0.42 2.736 2.831
25 0.970 0.951 0.96 0.52 2.984 3.079
26 0.996 0.983 0.94 0.61 3.191 3.288
27 1.060 1.048 0.99 0.73 3.368 3.462
28 1.000 1.000 1.00 1.00 4.183 4.310
-- resample ablation
site disruption (median) retained along KL dloss pullback
0 nan 0.000 nan nan 0.000 0.000 1.00
1 0.058 0.004 0.42 0.00 0.079 0.069 0.98
2 0.101 0.011 0.52 0.00 0.157 0.155 0.96
3 0.145 0.025 0.62 0.00 0.253 0.250 0.94
4 0.176 0.042 0.64 0.01 0.317 0.320 0.92
5 0.217 0.070 0.85 0.01 0.416 0.426 0.91
6 0.263 0.113 0.88 0.01 0.529 0.542 0.88
7 0.282 0.139 0.93 0.01 0.564 0.579 0.87
8 0.319 0.184 0.92 0.02 0.654 0.674 0.85
9 0.334 0.200 0.97 0.02 0.688 0.710 0.84
10 0.371 0.237 1.05 0.02 0.777 0.806 0.81
11 0.407 0.282 1.04 0.03 0.887 0.925 0.79
12 0.436 0.316 1.02 0.03 0.976 1.011 0.77
13 0.476 0.359 1.11 0.04 1.102 1.138 0.74
14 0.514 0.395 1.14 0.04 1.243 1.292 0.72
15 0.590 0.476 1.31 0.05 1.500 1.563 0.67
16 0.743 0.653 1.40 0.06 2.075 2.174 0.59
17 0.842 0.753 1.34 0.08 2.446 2.562 0.54
18 0.996 0.930 1.35 0.11 3.084 3.210 0.45
19 1.172 1.111 1.37 0.15 3.704 3.843 0.36
20 1.289 1.230 1.30 0.20 4.206 4.347 0.30
21 1.411 1.347 1.22 0.27 4.689 4.831 0.24
22 1.484 1.425 1.13 0.34 4.975 5.114 0.20
23 1.572 1.511 1.06 0.41 5.227 5.373 0.17
24 1.632 1.570 0.98 0.48 5.466 5.611 0.14
25 1.722 1.663 0.94 0.56 5.728 5.866 0.11
26 1.821 1.774 0.93 0.65 6.033 6.164 0.07
27 1.927 1.897 0.98 0.80 6.284 6.402 0.03
28 2.015 1.998 1.00 1.00 6.463 6.583 0.00
-- global ablation
site disruption (median) retained along KL dloss
0 1.079 1.033 nan 0.01 3.384 3.399
1 1.195 1.190 5.10 0.02 3.437 3.451
2 1.261 1.256 4.57 0.02 3.701 3.711
3 1.068 1.024 3.63 0.02 3.377 3.403
4 1.038 0.991 2.99 0.02 3.413 3.428
5 1.029 0.976 2.95 0.02 3.335 3.341
6 1.033 0.987 2.49 0.02 3.313 3.325
7 1.043 0.989 2.46 0.03 3.289 3.303
8 1.085 1.040 2.38 0.03 3.582 3.604
9 1.030 0.970 2.23 0.03 3.419 3.425
10 1.034 0.973 2.23 0.03 3.495 3.517
11 0.997 0.926 1.95 0.03 3.528 3.528
12 1.046 0.977 1.71 0.04 3.583 3.587
13 1.021 0.948 1.62 0.04 3.432 3.448
14 1.042 0.970 1.62 0.04 3.575 3.602
15 1.047 0.960 1.61 0.04 3.588 3.616
16 1.053 0.982 1.70 0.05 3.465 3.488
17 1.186 1.128 1.51 0.07 3.644 3.659
18 1.197 1.125 1.48 0.10 3.892 3.905
19 1.246 1.158 1.44 0.13 4.159 4.192
20 1.276 1.183 1.35 0.18 4.440 4.480
21 1.302 1.216 1.24 0.24 4.750 4.797
22 1.293 1.201 1.14 0.31 4.835 4.887
23 1.359 1.275 1.08 0.37 5.064 5.136
24 1.384 1.304 1.01 0.44 5.196 5.272
25 1.387 1.297 0.97 0.53 5.402 5.478
26 1.393 1.307 0.95 0.62 5.491 5.564
27 1.466 1.379 1.00 0.72 5.501 5.569
28 1.349 1.308 1.00 1.00 6.551 6.649
== Qwen3-1.7B-Base (3000 samples, float32; values after the last block)
max |diff|^2 at sites before the patch site: 0 (must be 0)
mean-ablation disruption first reaches 0.5 at site 18 (relative depth 0.64)
-- mean ablation
site disruption (median) retained along KL dloss
0 nan 0.000 nan nan 0.000 -0.000
1 0.028 0.002 0.61 0.00 0.032 0.033
2 0.053 0.005 0.65 0.00 0.063 0.061
3 0.074 0.009 0.80 0.00 0.093 0.084
4 0.091 0.015 0.83 0.00 0.116 0.106
5 0.110 0.024 1.03 0.00 0.143 0.130
6 0.131 0.040 1.07 0.00 0.184 0.173
7 0.148 0.051 1.12 0.01 0.204 0.189
8 0.162 0.063 1.09 0.01 0.224 0.207
9 0.174 0.074 1.13 0.01 0.235 0.222
10 0.196 0.093 1.19 0.01 0.264 0.250
11 0.213 0.115 1.10 0.01 0.283 0.268
12 0.223 0.126 1.04 0.01 0.295 0.279
13 0.233 0.138 1.04 0.01 0.313 0.294
14 0.247 0.150 1.01 0.02 0.329 0.310
15 0.282 0.188 1.12 0.02 0.378 0.354
16 0.375 0.274 1.21 0.03 0.533 0.501
17 0.442 0.345 1.24 0.04 0.663 0.631
18 0.594 0.502 1.28 0.06 1.015 0.984
19 0.767 0.690 1.33 0.08 1.449 1.440
20 0.888 0.821 1.25 0.12 1.858 1.865
21 1.023 0.954 1.18 0.16 2.285 2.313
22 1.076 1.014 1.09 0.20 2.504 2.535
23 1.135 1.083 1.02 0.25 2.846 2.888
24 1.169 1.122 0.95 0.31 3.124 3.164
25 1.196 1.143 0.93 0.38 3.417 3.454
26 1.196 1.154 0.93 0.47 3.640 3.674
27 1.225 1.229 0.96 0.58 3.589 3.614
28 1.000 1.000 1.00 1.00 4.484 4.533
-- resample ablation
site disruption (median) retained along KL dloss pullback
0 nan 0.000 nan nan 0.000 0.000 1.00
1 0.059 0.004 0.82 0.00 0.077 0.077 0.98
2 0.105 0.011 1.03 0.00 0.157 0.154 0.96
3 0.151 0.025 1.19 0.00 0.233 0.226 0.94
4 0.182 0.042 1.17 0.00 0.290 0.281 0.92
5 0.221 0.068 1.42 0.00 0.364 0.358 0.91
6 0.268 0.116 1.46 0.01 0.458 0.457 0.89
7 0.288 0.136 1.50 0.01 0.500 0.490 0.88
8 0.323 0.180 1.52 0.01 0.576 0.564 0.86
9 0.337 0.196 1.51 0.01 0.605 0.595 0.85
10 0.376 0.233 1.58 0.01 0.693 0.685 0.82
11 0.420 0.283 1.52 0.02 0.818 0.810 0.80
12 0.461 0.325 1.50 0.02 0.927 0.928 0.77
13 0.505 0.375 1.57 0.03 1.054 1.056 0.74
14 0.554 0.429 1.60 0.03 1.213 1.227 0.71
15 0.639 0.516 1.70 0.04 1.482 1.499 0.66
16 0.828 0.733 1.77 0.05 2.187 2.225 0.57
17 0.942 0.861 1.61 0.06 2.650 2.702 0.51
18 1.102 1.032 1.54 0.09 3.436 3.501 0.43
19 1.275 1.202 1.48 0.12 4.144 4.224 0.33
20 1.378 1.312 1.31 0.16 4.692 4.770 0.28
21 1.515 1.447 1.20 0.21 5.293 5.368 0.20
22 1.585 1.517 1.10 0.27 5.570 5.644 0.17
23 1.682 1.609 1.02 0.33 5.876 5.954 0.13
24 1.734 1.660 0.95 0.39 6.157 6.230 0.11
25 1.816 1.752 0.92 0.48 6.451 6.515 0.07
26 1.882 1.825 0.92 0.58 6.714 6.771 0.05
27 1.957 1.913 0.96 0.71 6.959 7.007 0.02
28 2.018 1.981 1.00 1.00 7.156 7.201 0.00
-- global ablation
site disruption (median) retained along KL dloss
0 1.130 1.056 nan 0.01 3.467 3.455
1 1.326 1.332 11.76 0.01 3.869 3.836
2 1.275 1.258 9.11 0.01 3.808 3.752
3 1.110 1.029 6.95 0.01 3.524 3.484
4 1.088 0.984 5.90 0.01 3.640 3.600
5 1.068 0.989 5.60 0.01 3.564 3.530
6 1.077 0.996 4.82 0.01 3.583 3.544
7 1.052 0.961 4.49 0.02 3.434 3.404
8 1.065 0.969 4.02 0.01 3.553 3.524
9 1.084 1.004 3.82 0.02 3.695 3.693
10 1.110 1.018 3.60 0.02 3.576 3.567
11 1.154 1.069 3.07 0.02 3.855 3.825
12 1.148 1.077 2.55 0.02 3.785 3.770
13 1.115 1.033 2.40 0.02 3.753 3.737
14 1.096 1.005 2.15 0.02 3.630 3.627
15 1.084 0.982 2.12 0.03 3.677 3.676
16 1.093 0.982 2.04 0.03 3.743 3.726
17 1.336 1.254 1.82 0.04 3.902 3.872
18 1.480 1.445 1.72 0.06 3.856 3.821
19 1.439 1.328 1.59 0.09 4.423 4.393
20 1.547 1.456 1.41 0.12 4.889 4.871
21 1.622 1.528 1.26 0.16 5.389 5.383
22 1.614 1.510 1.13 0.20 5.480 5.481
23 1.636 1.534 1.04 0.25 5.916 5.929
24 1.642 1.543 0.98 0.30 6.060 6.076
25 1.639 1.501 0.95 0.38 6.170 6.184
26 1.634 1.474 0.96 0.47 6.182 6.190
27 1.639 1.551 0.97 0.55 5.755 5.755
28 1.323 1.289 1.00 1.00 6.734 6.757
== Qwen3-4B-Base (3000 samples, float32; values after the last block)
max |diff|^2 at sites before the patch site: 0 (must be 0)
mean-ablation disruption first reaches 0.5 at site 23 (relative depth 0.64)
-- mean ablation
site disruption (median) retained along KL dloss
0 nan 0.000 nan nan 0.000 -0.000
1 0.016 0.000 0.21 0.00 0.017 0.014
2 0.032 0.001 0.09 0.00 0.039 0.034
3 0.039 0.002 0.10 0.00 0.047 0.040
4 0.047 0.004 0.12 0.00 0.056 0.051
5 0.057 0.006 0.15 0.00 0.069 0.061
6 0.084 0.012 0.20 0.00 0.106 0.104
7 0.107 0.022 0.25 0.00 0.142 0.136
8 0.119 0.032 0.24 0.00 0.164 0.156
9 0.122 0.037 0.23 0.00 0.173 0.165
10 0.133 0.044 0.25 0.01 0.182 0.172
11 0.139 0.052 0.31 0.01 0.186 0.173
12 0.143 0.058 0.32 0.01 0.190 0.176
13 0.153 0.064 0.36 0.01 0.210 0.200
14 0.164 0.070 0.40 0.01 0.221 0.212
15 0.173 0.076 0.43 0.01 0.231 0.221
16 0.185 0.084 0.47 0.01 0.244 0.231
17 0.203 0.092 0.32 0.02 0.260 0.248
18 0.216 0.104 0.38 0.02 0.275 0.264
19 0.231 0.117 0.42 0.02 0.305 0.297
20 0.277 0.152 0.48 0.03 0.399 0.387
21 0.332 0.202 0.56 0.03 0.487 0.467
22 0.423 0.303 0.72 0.03 0.638 0.615
23 0.552 0.460 0.96 0.05 0.866 0.844
24 0.661 0.585 1.11 0.07 1.091 1.076
25 0.750 0.677 1.16 0.10 1.312 1.306
26 0.835 0.771 1.21 0.12 1.615 1.619
27 0.901 0.848 1.23 0.14 1.786 1.800
28 0.987 0.939 1.23 0.16 2.045 2.067
29 1.069 1.024 1.20 0.19 2.431 2.463
30 1.114 1.073 1.14 0.22 2.645 2.674
31 1.166 1.138 1.05 0.27 2.973 3.010
32 1.185 1.162 1.01 0.32 3.270 3.307
33 1.112 1.084 1.00 0.40 3.559 3.601
34 1.071 1.013 0.97 0.51 3.735 3.773
35 1.114 1.106 1.00 0.62 3.892 3.935
36 1.000 1.000 1.00 1.00 4.902 4.966
-- resample ablation
site disruption (median) retained along KL dloss pullback
0 nan 0.000 nan nan 0.000 0.000 1.00
1 0.043 0.001 0.34 0.00 0.057 0.054 0.98
2 0.072 0.003 0.16 0.00 0.110 0.108 0.97
3 0.078 0.006 0.16 0.00 0.120 0.115 0.97
4 0.103 0.011 0.20 0.00 0.157 0.150 0.96
5 0.125 0.019 0.24 0.00 0.195 0.191 0.95
6 0.161 0.036 0.29 0.00 0.259 0.257 0.93
7 0.194 0.062 0.34 0.01 0.321 0.322 0.92
8 0.219 0.084 0.33 0.01 0.370 0.370 0.91
9 0.234 0.097 0.31 0.01 0.396 0.398 0.90
10 0.243 0.108 0.33 0.01 0.410 0.411 0.90
11 0.258 0.123 0.39 0.01 0.440 0.442 0.89
12 0.269 0.134 0.40 0.01 0.459 0.461 0.88
13 0.291 0.150 0.45 0.01 0.509 0.513 0.87
14 0.309 0.165 0.49 0.01 0.555 0.562 0.86
15 0.339 0.190 0.57 0.01 0.633 0.637 0.85
16 0.365 0.210 0.63 0.02 0.707 0.710 0.83
17 0.391 0.236 0.14 0.02 0.780 0.781 0.81
18 0.432 0.272 0.16 0.02 0.894 0.900 0.78
19 0.519 0.362 0.20 0.03 1.172 1.185 0.73
20 0.663 0.533 0.26 0.03 1.677 1.701 0.65
21 0.749 0.640 0.31 0.04 1.992 2.028 0.61
22 0.856 0.770 0.36 0.04 2.437 2.488 0.55
23 1.016 0.944 0.51 0.07 3.089 3.144 0.46
24 1.167 1.116 0.69 0.10 3.879 3.934 0.37
25 1.272 1.230 0.82 0.13 4.262 4.317 0.31
26 1.345 1.301 0.90 0.16 4.772 4.830 0.27
27 1.389 1.346 0.96 0.19 5.013 5.073 0.25
28 1.444 1.404 0.99 0.22 5.286 5.348 0.22
29 1.500 1.458 1.00 0.26 5.598 5.666 0.20
30 1.528 1.487 0.99 0.31 5.759 5.818 0.18
31 1.638 1.611 0.96 0.36 6.156 6.221 0.14
32 1.688 1.660 0.94 0.42 6.405 6.465 0.12
33 1.746 1.713 0.93 0.49 6.643 6.705 0.09
34 1.790 1.765 0.95 0.58 6.857 6.915 0.07
35 1.925 1.902 1.00 0.71 7.275 7.330 0.02
36 2.007 1.984 1.00 1.00 7.563 7.615 0.00
-- global ablation
site disruption (median) retained along KL dloss
0 1.288 1.278 nan 0.01 3.623 3.652
1 1.348 1.321 5.11 0.01 4.039 4.062
2 1.262 1.200 2.71 0.01 3.893 3.895
3 1.246 1.189 2.34 0.01 3.964 3.962
4 1.169 1.106 2.07 0.01 3.666 3.657
5 1.131 1.066 1.88 0.01 3.554 3.550
6 1.177 1.125 1.62 0.01 3.637 3.631
7 1.160 1.098 1.44 0.01 3.542 3.550
8 1.152 1.078 1.28 0.01 3.704 3.716
9 1.139 1.054 1.18 0.01 3.728 3.732
10 1.172 1.092 1.23 0.01 3.824 3.828
11 1.168 1.089 1.32 0.01 3.786 3.789
12 1.134 1.065 1.30 0.01 3.750 3.754
13 1.175 1.102 1.35 0.01 3.988 3.983
14 1.212 1.156 1.35 0.01 4.067 4.068
15 1.162 1.093 1.35 0.02 3.882 3.873
16 1.249 1.175 1.45 0.02 4.258 4.245
17 1.165 1.094 1.23 0.01 3.989 3.983
18 1.257 1.195 1.34 0.02 4.179 4.164
19 1.226 1.156 1.33 0.02 4.158 4.136
20 1.243 1.161 1.29 0.02 4.262 4.237
21 1.266 1.156 1.39 0.03 4.191 4.172
22 1.361 1.270 1.59 0.04 3.982 3.974
23 1.443 1.377 1.74 0.05 4.064 4.041
24 1.511 1.464 1.71 0.07 4.388 4.365
25 1.497 1.436 1.60 0.10 4.462 4.445
26 1.473 1.397 1.54 0.12 4.848 4.844
27 1.512 1.432 1.48 0.14 5.027 5.028
28 1.595 1.522 1.42 0.16 5.395 5.403
29 1.638 1.573 1.34 0.20 5.865 5.874
30 1.662 1.577 1.25 0.22 6.207 6.228
31 1.674 1.593 1.13 0.27 6.375 6.402
32 1.681 1.587 1.07 0.31 6.618 6.645
33 1.526 1.425 1.04 0.40 6.766 6.798
34 1.433 1.337 1.03 0.50 6.713 6.742
35 1.549 1.515 1.02 0.58 6.226 6.254
36 1.371 1.342 1.00 1.00 7.331 7.384
== Qwen3-8B-Base (3000 samples, float32; values after the last block)
max |diff|^2 at sites before the patch site: 0 (must be 0)
mean-ablation disruption first reaches 0.5 at site 22 (relative depth 0.61)
-- mean ablation
site disruption (median) retained along KL dloss
0 nan 0.000 nan nan -0.000 0.000
1 0.015 0.001 0.33 0.00 0.015 0.016
2 0.033 0.002 0.38 0.00 0.028 0.025
3 0.041 0.002 0.32 0.00 0.035 0.033
4 0.050 0.004 0.36 0.00 0.045 0.043
5 0.063 0.007 0.36 0.00 0.058 0.055
6 0.092 0.013 0.47 0.00 0.095 0.096
7 0.112 0.022 0.46 0.00 0.122 0.127
8 0.124 0.032 0.45 0.00 0.138 0.145
9 0.129 0.038 0.41 0.01 0.149 0.157
10 0.140 0.046 0.44 0.01 0.159 0.161
11 0.148 0.055 0.52 0.01 0.168 0.165
12 0.154 0.061 0.55 0.01 0.173 0.173
13 0.168 0.072 0.59 0.01 0.198 0.199
14 0.180 0.080 0.66 0.01 0.210 0.213
15 0.196 0.090 0.69 0.01 0.226 0.230
16 0.208 0.101 0.73 0.01 0.237 0.242
17 0.226 0.110 0.06 0.02 0.250 0.260
18 0.239 0.125 0.07 0.02 0.264 0.272
19 0.260 0.141 0.09 0.02 0.300 0.307
20 0.318 0.183 0.12 0.03 0.402 0.407
21 0.392 0.248 0.16 0.03 0.495 0.495
22 0.504 0.384 0.23 0.03 0.633 0.624
23 0.661 0.580 0.35 0.04 0.848 0.839
24 0.797 0.729 0.51 0.06 1.091 1.094
25 0.916 0.848 0.65 0.08 1.325 1.337
26 1.048 0.995 0.76 0.09 1.642 1.661
27 1.128 1.089 0.83 0.11 1.824 1.851
28 1.227 1.200 0.90 0.13 2.118 2.156
29 1.315 1.306 0.93 0.15 2.532 2.587
30 1.326 1.322 0.93 0.18 2.750 2.800
31 1.338 1.347 0.91 0.22 3.137 3.195
32 1.406 1.420 0.91 0.25 3.464 3.522
33 1.307 1.301 0.91 0.32 3.800 3.860
34 1.247 1.222 0.93 0.40 3.846 3.902
35 1.168 1.182 0.99 0.53 3.969 4.032
36 1.000 1.000 1.00 1.00 4.889 4.973
-- resample ablation
site disruption (median) retained along KL dloss pullback
0 nan 0.000 nan nan 0.000 0.000 1.00
1 0.045 0.001 0.49 0.00 0.047 0.048 0.98
2 0.074 0.004 0.59 0.00 0.093 0.098 0.97
3 0.084 0.006 0.50 0.00 0.103 0.107 0.97
4 0.110 0.011 0.55 0.00 0.138 0.141 0.95
5 0.131 0.019 0.56 0.00 0.177 0.183 0.95
6 0.173 0.039 0.63 0.01 0.235 0.245 0.93
7 0.210 0.064 0.61 0.01 0.294 0.305 0.92
8 0.231 0.087 0.57 0.01 0.331 0.344 0.91
9 0.246 0.103 0.53 0.01 0.361 0.373 0.90
10 0.258 0.119 0.56 0.01 0.378 0.392 0.90
11 0.282 0.137 0.63 0.01 0.407 0.420 0.89
12 0.300 0.152 0.65 0.01 0.437 0.451 0.88
13 0.331 0.174 0.72 0.01 0.507 0.514 0.86
14 0.348 0.185 0.77 0.01 0.552 0.561 0.85
15 0.389 0.222 0.87 0.02 0.656 0.665 0.83
16 0.421 0.251 0.93 0.02 0.738 0.749 0.81
17 0.453 0.283 0.17 0.02 0.809 0.819 0.79
18 0.502 0.331 0.20 0.02 0.938 0.954 0.76
19 0.602 0.451 0.25 0.03 1.233 1.252 0.70
20 0.764 0.644 0.34 0.03 1.762 1.792 0.62
21 0.851 0.752 0.40 0.04 2.071 2.112 0.58
22 0.957 0.880 0.48 0.05 2.529 2.590 0.53
23 1.115 1.064 0.65 0.06 3.233 3.289 0.44
24 1.244 1.197 0.84 0.09 4.078 4.132 0.35
25 1.340 1.304 0.97 0.12 4.506 4.563 0.29
26 1.407 1.370 1.03 0.15 5.046 5.108 0.25
27 1.450 1.418 1.06 0.17 5.314 5.378 0.23
28 1.500 1.479 1.07 0.20 5.604 5.671 0.20
29 1.553 1.526 1.05 0.24 5.947 6.025 0.18
30 1.581 1.559 1.01 0.28 6.116 6.190 0.16
31 1.677 1.662 0.97 0.33 6.521 6.599 0.12
32 1.723 1.715 0.94 0.38 6.791 6.860 0.10
33 1.784 1.761 0.94 0.44 7.035 7.101 0.08
34 1.819 1.797 0.95 0.52 7.235 7.296 0.06
35 1.937 1.907 0.99 0.64 7.664 7.725 0.02
36 2.001 1.977 1.00 1.00 7.962 8.022 0.00
-- global ablation
site disruption (median) retained along KL dloss
0 1.544 1.607 nan 0.00 3.834 3.848
1 1.432 1.391 9.10 0.01 3.890 3.892
2 1.388 1.337 6.95 0.01 3.790 3.808
3 1.386 1.346 5.77 0.01 3.915 3.914
4 1.336 1.285 5.09 0.01 3.686 3.688
5 1.305 1.253 4.35 0.01 3.591 3.592
6 1.349 1.317 3.54 0.01 3.678 3.669
7 1.316 1.273 2.77 0.01 3.579 3.569
8 1.299 1.228 2.47 0.01 3.686 3.675
9 1.319 1.249 2.26 0.01 3.766 3.764
10 1.340 1.284 2.27 0.01 3.790 3.790
11 1.297 1.221 2.28 0.02 3.864 3.883
12 1.343 1.283 2.32 0.02 3.963 3.975
13 1.337 1.279 2.27 0.02 3.980 3.983
14 1.385 1.331 2.20 0.02 4.096 4.101
15 1.336 1.270 2.18 0.02 3.980 3.979
16 1.430 1.372 2.17 0.02 4.176 4.183
17 1.402 1.345 0.28 0.02 4.056 4.054
18 1.426 1.363 0.31 0.02 3.980 3.991
19 1.452 1.398 0.33 0.02 4.058 4.063
20 1.489 1.431 0.37 0.02 4.260 4.262
21 1.614 1.562 0.43 0.03 4.232 4.235
22 1.647 1.621 0.51 0.03 3.903 3.898
23 1.774 1.782 0.65 0.04 4.009 3.999
24 1.858 1.886 0.82 0.06 4.436 4.429
25 1.823 1.830 0.92 0.08 4.589 4.590
26 1.899 1.883 1.00 0.09 5.048 5.053
27 1.959 1.950 1.03 0.11 5.289 5.300
28 2.037 2.032 1.06 0.12 5.704 5.725
29 2.059 2.042 1.06 0.14 6.178 6.203
30 2.019 1.970 1.03 0.17 6.430 6.460
31 1.956 1.897 0.98 0.21 6.650 6.686
32 2.067 1.979 0.96 0.24 6.909 6.947
33 1.886 1.783 0.96 0.29 7.080 7.123
34 1.790 1.637 0.96 0.36 6.684 6.722
35 1.610 1.533 1.00 0.49 6.376 6.421
36 1.379 1.333 1.00 1.00 7.263 7.329
== Qwen3-14B-Base (3000 samples, float32; values after the last block)
max |diff|^2 at sites before the patch site: 0 (must be 0)
mean-ablation disruption first reaches 0.5 at site 26 (relative depth 0.65)
-- mean ablation
site disruption (median) retained along KL dloss
0 nan 0.000 nan nan -0.000 -0.000
1 0.009 0.000 0.23 0.00 0.012 0.014
2 0.024 0.001 0.29 0.00 0.025 0.023
3 0.032 0.002 0.31 0.00 0.036 0.033
4 0.037 0.002 0.30 0.00 0.041 0.038
5 0.043 0.003 0.30 0.00 0.045 0.038
6 0.050 0.005 0.31 0.00 0.052 0.043
7 0.062 0.007 0.33 0.00 0.062 0.052
8 0.070 0.010 0.32 0.00 0.074 0.063
9 0.089 0.015 0.35 0.00 0.096 0.082
10 0.100 0.020 0.41 0.00 0.110 0.097
11 0.114 0.028 0.40 0.00 0.133 0.114
12 0.119 0.035 0.36 0.00 0.145 0.128
13 0.123 0.038 0.37 0.00 0.151 0.136
14 0.129 0.044 0.40 0.00 0.155 0.136
15 0.137 0.050 0.45 0.00 0.166 0.148
16 0.145 0.057 0.49 0.00 0.172 0.154
17 0.169 0.067 0.54 0.01 0.207 0.186
18 0.191 0.080 0.59 0.01 0.234 0.217
19 0.205 0.087 0.65 0.01 0.251 0.229
20 0.229 0.103 0.63 0.01 0.284 0.262
21 0.255 0.126 0.73 0.01 0.323 0.299
22 0.297 0.160 0.79 0.01 0.376 0.346
23 0.336 0.199 0.87 0.01 0.417 0.383
24 0.367 0.230 0.93 0.02 0.450 0.421
25 0.420 0.290 1.04 0.02 0.501 0.475
26 0.544 0.418 1.20 0.02 0.635 0.612
27 0.663 0.555 1.25 0.03 0.788 0.764
28 0.836 0.759 1.31 0.04 1.112 1.087
29 0.977 0.922 1.31 0.04 1.369 1.346
30 1.153 1.122 1.30 0.05 1.754 1.745
31 1.288 1.245 1.27 0.06 2.181 2.180
32 1.339 1.278 1.20 0.08 2.468 2.481
33 1.328 1.238 1.13 0.09 2.749 2.768
34 1.250 1.168 1.07 0.12 3.018 3.034
35 1.216 1.137 1.03 0.14 3.172 3.188
36 1.598 1.513 1.01 0.13 3.518 3.537
37 1.605 1.526 1.00 0.16 3.759 3.789
38 1.121 1.075 1.02 0.26 4.082 4.112
39 1.092 1.093 1.10 0.44 4.382 4.413
40 1.000 1.000 1.00 1.00 6.052 6.106
-- resample ablation
site disruption (median) retained along KL dloss pullback
0 nan 0.000 nan nan 0.000 0.000 1.00
1 0.030 0.001 0.39 0.00 0.039 0.042 0.99
2 0.060 0.002 0.47 0.00 0.086 0.087 0.98
3 0.075 0.004 0.47 0.00 0.103 0.098 0.97
4 0.086 0.006 0.48 0.00 0.120 0.115 0.97
5 0.098 0.008 0.50 0.00 0.132 0.128 0.96
6 0.117 0.013 0.46 0.00 0.159 0.151 0.95
7 0.131 0.019 0.46 0.00 0.182 0.170 0.95
8 0.146 0.028 0.44 0.00 0.206 0.191 0.94
9 0.180 0.049 0.45 0.00 0.253 0.238 0.93
10 0.201 0.061 0.48 0.00 0.281 0.270 0.92
11 0.221 0.077 0.47 0.00 0.317 0.303 0.91
12 0.241 0.093 0.45 0.00 0.347 0.330 0.91
13 0.251 0.101 0.45 0.00 0.358 0.342 0.90
14 0.266 0.111 0.48 0.01 0.375 0.361 0.90
15 0.292 0.130 0.53 0.01 0.424 0.410 0.89
16 0.313 0.146 0.57 0.01 0.466 0.454 0.88
17 0.369 0.189 0.66 0.01 0.591 0.585 0.85
18 0.422 0.235 0.76 0.01 0.716 0.705 0.82
19 0.457 0.269 0.85 0.01 0.809 0.797 0.80
20 0.500 0.314 0.91 0.01 0.927 0.917 0.77
21 0.565 0.384 1.00 0.01 1.077 1.069 0.73
22 0.675 0.501 1.13 0.01 1.370 1.372 0.67
23 0.806 0.662 1.25 0.02 1.793 1.802 0.60
24 0.887 0.766 1.30 0.02 2.067 2.076 0.56
25 1.006 0.908 1.45 0.03 2.487 2.489 0.49
26 1.135 1.052 1.54 0.03 3.022 3.016 0.41
27 1.245 1.179 1.52 0.04 3.609 3.604 0.35
28 1.346 1.281 1.47 0.05 4.258 4.257 0.30
29 1.404 1.335 1.43 0.06 4.677 4.675 0.27
30 1.493 1.440 1.35 0.08 5.188 5.194 0.22
31 1.560 1.509 1.28 0.10 5.702 5.720 0.19
32 1.628 1.592 1.21 0.12 6.052 6.077 0.16
33 1.688 1.651 1.13 0.14 6.381 6.406 0.14
34 1.718 1.691 1.07 0.16 6.588 6.607 0.12
35 1.746 1.715 1.03 0.18 6.687 6.702 0.11
36 1.782 1.755 1.00 0.21 6.925 6.937 0.10
37 1.832 1.800 0.99 0.24 7.170 7.185 0.07
38 1.916 1.857 1.01 0.30 7.495 7.501 0.04
39 1.988 1.927 1.11 0.49 7.840 7.841 0.01
40 2.035 1.990 1.00 1.00 8.208 8.211 0.00
-- global ablation
site disruption (median) retained along KL dloss
0 1.460 1.381 nan 0.00 3.866 3.853
1 1.397 1.348 8.81 0.01 3.999 3.976
2 1.369 1.309 6.08 0.01 3.929 3.924
3 1.337 1.295 5.13 0.01 3.936 3.927
4 1.319 1.275 4.44 0.01 3.909 3.888
5 1.366 1.316 3.81 0.01 3.890 3.881
6 1.335 1.301 3.10 0.01 3.959 3.952
7 1.307 1.264 2.74 0.01 4.052 4.021
8 1.305 1.253 2.74 0.01 4.059 4.039
9 1.314 1.268 2.44 0.01 4.016 3.994
10 1.332 1.280 2.08 0.01 4.031 4.015
11 1.354 1.292 1.89 0.01 4.165 4.132
12 1.354 1.278 1.77 0.01 4.050 4.048
13 1.366 1.315 1.77 0.01 4.188 4.171
14 1.363 1.305 1.81 0.01 4.284 4.268
15 1.347 1.268 1.78 0.01 4.054 4.066
16 1.332 1.257 1.81 0.01 3.992 3.996
17 1.321 1.231 1.77 0.01 4.055 4.015
18 1.336 1.253 1.68 0.01 4.016 3.996
19 1.378 1.305 1.71 0.01 3.947 3.910
20 1.320 1.228 1.45 0.01 3.783 3.771
21 1.343 1.252 1.42 0.01 3.842 3.823
22 1.409 1.340 1.55 0.01 3.861 3.841
23 1.441 1.388 1.52 0.01 4.048 4.030
24 1.434 1.387 1.44 0.01 3.923 3.910
25 1.470 1.443 1.60 0.02 3.915 3.905
26 1.572 1.560 1.70 0.02 4.145 4.129
27 1.619 1.598 1.62 0.03 4.282 4.266
28 1.756 1.787 1.56 0.03 4.732 4.712
29 1.884 1.957 1.50 0.04 5.170 5.159
30 1.968 2.047 1.42 0.05 5.553 5.547
31 1.801 1.837 1.33 0.06 5.920 5.907
32 1.626 1.660 1.22 0.08 6.260 6.246
33 1.529 1.554 1.13 0.10 6.369 6.366
34 1.557 1.575 1.07 0.11 6.639 6.644
35 1.466 1.473 1.03 0.14 6.648 6.658
36 2.145 2.227 1.02 0.13 7.163 7.188
37 2.008 1.995 1.02 0.17 6.767 6.784
38 1.586 1.593 1.00 0.22 6.859 6.873
39 1.536 1.538 1.12 0.40 7.126 7.149
40 1.409 1.416 1.00 1.00 9.252 9.301
== 00:53:38 done
== 22:12:47 install
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning.
== 22:13:10 download data
[90mHint: The `hf-cli` skill is not installed. Run `hf skills add -g` to teach your AI agents how to use the `hf` CLI.[0m
Fetching 10 files: 0%| | 0/10 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 10 files: 20%|██ | 2/10 [00:00<00:00, 14.55it/s]Fetching 10 files: 50%|█████ | 5/10 [00:00<00:00, 20.71it/s]Fetching 10 files: 80%|████████ | 8/10 [00:00<00:00, 17.68it/s][32m✓ Downloaded[0m
path: /workspace/act-experiments/data/dl/sample/10BT/000_00000.parquet
== 22:13:19 test
Fetching 10 files: 100%|██████████| 10/10 [00:10<00:00, 1.69s/it]Fetching 10 files: 100%|██████████| 10/10 [00:10<00:00, 1.08s/it]
Fetching 10 files: 0%| | 0/10 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 10 files: 10%|█ | 1/10 [00:00<00:01, 6.55it/s]Fetching 10 files: 40%|████ | 4/10 [00:00<00:00, 17.86it/s]Fetching 10 files: 70%|███████ | 7/10 [00:00<00:00, 21.51it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 19%|█▊ | 58/310 [00:00<00:00, 576.47it/s]Loading weights: 37%|███▋ | 116/310 [00:00<00:00, 564.78it/s]Loading weights: 56%|█████▌ | 173/310 [00:00<00:00, 404.70it/s]Loading weights: 70%|███████ | 218/310 [00:00<00:00, 318.40it/s]Loading weights: 82%|████████▏ | 254/310 [00:00<00:00, 320.11it/s]Loading weights: 93%|█████████▎| 289/310 [00:00<00:00, 262.24it/s]Loading weights: 100%|██████████| 310/310 [00:01<00:00, 309.90it/s]
patch site -1: max relative error resid 1.7e-06
patch site -1: max relative error attn 2.9e-06
patch site -1: max relative error mlp 3.2e-06
patch site -1: max relative error logits 1.3e-06
patch site 0: max relative error resid 2.4e-06
patch site 0: max relative error attn 3.9e-06
patch site 0: max relative error mlp 3.9e-06
patch site 0: max relative error logits 4.2e-06
patch site 1: max relative error resid 2.2e-06
patch site 1: max relative error attn 4.3e-06
patch site 1: max relative error mlp 3.0e-06
patch site 1: max relative error logits 1.4e-06
patch site 14: max relative error resid 5.3e-06
patch site 14: max relative error attn 7.8e-06
patch site 14: max relative error mlp 8.6e-06
patch site 14: max relative error logits 6.0e-06
patch site 28: max relative error resid 1.7e-06
patch site 28: max relative error attn 2.9e-06
patch site 28: max relative error mlp 3.2e-06
patch site 28: max relative error logits 3.4e-07
OK
[W1014 22:13:31.519030231 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:13:31 means 0.6B
Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 100%|██████████| 310/310 [00:00<00:00, 8214.72it/s]
Fetching 10 files: 100%|██████████| 10/10 [00:14<00:00, 2.02s/it]Fetching 10 files: 100%|██████████| 10/10 [00:14<00:00, 1.42s/it]
Fetching 13 files: 0%| | 0/13 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 13 files: 15%|█▌ | 2/13 [00:00<00:00, 13.06it/s]Fetching 13 files: 31%|███ | 4/13 [00:00<00:00, 15.96it/s]Fetching 13 files: 46%|████▌ | 6/13 [00:00<00:00, 16.99it/s]Fetching 13 files: 77%|███████▋ | 10/13 [00:00<00:00, 19.50it/s]Fetching 13 files: 92%|█████████▏| 12/13 [00:13<00:01, 1.93s/it]Fetching 13 files: 100%|██████████| 13/13 [00:22<00:00, 3.05s/it]Fetching 13 files: 100%|██████████| 13/13 [00:22<00:00, 1.71s/it]
Fetching 14 files: 0%| | 0/14 [00:00<?, ?it/s]Fetching 14 files: 7%|▋ | 1/14 [00:00<00:01, 9.99it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 14 files: 29%|██▊ | 4/14 [00:00<00:00, 14.95it/s]Fetching 14 files: 50%|█████ | 7/14 [00:00<00:00, 19.99it/s]34019 docs, 10000426 positions, 2139 tokens with >= 500 occurrences (73% of positions)
Fetching 14 files: 64%|██████▍ | 9/14 [00:20<00:00, 19.99it/s]6591/34019 docs
Fetching 14 files: 71%|███████▏ | 10/14 [00:25<00:14, 3.69s/it]10533/34019 docs
13510/34019 docs
Fetching 14 files: 79%|███████▊ | 11/14 [00:38<00:15, 5.25s/it]15912/34019 docs
17940/34019 docs
19703/34019 docs
Fetching 14 files: 86%|████████▌ | 12/14 [00:56<00:15, 7.74s/it]21310/34019 docs
Fetching 14 files: 93%|█████████▎| 13/14 [01:00<00:07, 7.02s/it]Fetching 14 files: 100%|██████████| 14/14 [01:02<00:00, 4.46s/it]
Fetching 18 files: 0%| | 0/18 [00:00<?, ?it/s]Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 18 files: 6%|▌ | 1/18 [00:00<00:02, 8.41it/s]Fetching 18 files: 28%|██▊ | 5/18 [00:00<00:01, 12.06it/s]22910/34019 docs
24510/34019 docs
26110/34019 docs
Fetching 18 files: 33%|███▎ | 6/18 [00:20<00:00, 12.06it/s]27710/34019 docs
29310/34019 docs
30910/34019 docs
32510/34019 docs
[W1014 22:15:48.788527690 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
== 22:15:48 patch 0.6B
Fetching 18 files: 39%|███▉ | 7/18 [00:47<01:38, 8.98s/it]Fetching 18 files: 44%|████▍ | 8/18 [00:47<01:11, 7.11s/it]Fetching 18 files: 50%|█████ | 9/18 [00:48<00:50, 5.60s/it]Fetching 18 files: 61%|██████ | 11/18 [00:48<00:24, 3.46s/it]Loading weights: 0%| | 0/310 [00:00<?, ?it/s]Loading weights: 24%|██▎ | 73/310 [00:00<00:00, 729.67it/s]Loading weights: 47%|████▋ | 146/310 [00:00<00:00, 466.36it/s]Loading weights: 64%|██████▍ | 198/310 [00:00<00:00, 470.29it/s]Loading weights: 80%|████████ | 249/310 [00:00<00:00, 367.45it/s]Loading weights: 94%|█████████▎| 290/310 [00:00<00:00, 296.28it/s]Loading weights: 100%|██████████| 310/310 [00:00<00:00, 346.15it/s]
Fetching 18 files: 67%|██████▋ | 12/18 [00:59<00:29, 4.94s/it]Fetching 18 files: 72%|███████▏ | 13/18 [01:09<00:31, 6.36s/it]Fetching 18 files: 78%|███████▊ | 14/18 [01:16<00:25, 6.29s/it]Fetching 18 files: 83%|████████▎ | 15/18 [01:19<00:16, 5.65s/it]100/3000 samples
Fetching 18 files: 89%|████████▉ | 16/18 [01:29<00:13, 6.66s/it]Fetching 18 files: 94%|█████████▍| 17/18 [01:36<00:06, 6.80s/it]200/3000 samples
Fetching 18 files: 100%|██████████| 18/18 [01:41<00:00, 6.31s/it]Fetching 18 files: 100%|██████████| 18/18 [01:41<00:00, 5.64s/it]
300/3000 samples
400/3000 samples
500/3000 samples
600/3000 samples
700/3000 samples
800/3000 samples
900/3000 samples
1000/3000 samples
1100/3000 samples
1200/3000 samples
1300/3000 samples
1400/3000 samples
1500/3000 samples
1600/3000 samples
1700/3000 samples
1800/3000 samples
1900/3000 samples
2000/3000 samples
2100/3000 samples
2200/3000 samples
2300/3000 samples
2400/3000 samples
2500/3000 samples
2600/3000 samples
2700/3000 samples
2800/3000 samples
2900/3000 samples
3000/3000 samples
[W1014 22:24:11.909273151 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
wandb: setting up run m39t7lyy
wandb: Tracking run with wandb version 0.30.0
wandb: Run data is saved locally in /workspace/act-experiments/wandb/run-20261014_222414-m39t7lyy
wandb: Run `wandb offline` to turn off syncing.
wandb: Syncing run Qwen3-0.6B-Base
wandb: ⭐️ View project at https://wandb.ai/octahedral-systems/qwen3-self-correction