Loading weights:   0%|          | 0/1951 [00:00<?, ?it/s]
Loading weights:   0%|          | 3/1951 [00:00<02:15, 14.41it/s]
Loading weights:   3%|3         | 60/1951 [00:00<00:07, 238.74it/s]
Loading weights:   6%|6         | 122/1951 [00:00<00:04, 376.08it/s]
Loading weights:   9%|9         | 182/1951 [00:00<00:03, 452.17it/s]
Loading weights:  12%|#1        | 234/1951 [00:00<00:03, 463.57it/s]
Loading weights:  15%|#4        | 285/1951 [00:00<00:03, 455.47it/s]
Loading weights:  18%|#7        | 344/1951 [00:00<00:03, 488.31it/s]
Loading weights:  21%|##1       | 410/1951 [00:00<00:02, 539.08it/s]
Loading weights:  24%|##3       | 466/1951 [00:01<00:02, 532.09it/s]
Loading weights:  27%|##6       | 521/1951 [00:01<00:02, 537.20it/s]
Loading weights:  30%|##9       | 578/1951 [00:01<00:02, 534.86it/s]
Loading weights:  33%|###2      | 640/1951 [00:01<00:02, 545.65it/s]
Loading weights:  36%|###5      | 702/1951 [00:01<00:02, 566.83it/s]
Loading weights:  39%|###8      | 760/1951 [00:11<01:04, 18.59it/s] 
Loading weights:  40%|###9      | 775/1951 [00:11<00:58, 20.26it/s]
Loading weights:  42%|####1     | 819/1951 [00:12<00:43, 26.01it/s]
Loading weights:  44%|####3     | 852/1951 [00:12<00:35, 31.19it/s]
Loading weights:  45%|####5     | 878/1951 [00:13<00:30, 34.98it/s]
Loading weights:  46%|####6     | 898/1951 [00:13<00:27, 38.80it/s]
Loading weights:  47%|####6     | 914/1951 [00:13<00:24, 42.81it/s]
Loading weights:  48%|####7     | 928/1951 [00:13<00:21, 46.79it/s]
Loading weights:  48%|####8     | 940/1951 [00:14<00:20, 48.81it/s]
Loading weights:  49%|####8     | 951/1951 [00:14<00:20, 49.62it/s]
Loading weights:  49%|####9     | 962/1951 [00:14<00:18, 54.77it/s]
Loading weights:  50%|####9     | 971/1951 [00:14<00:17, 56.88it/s]
Loading weights:  50%|#####     | 980/1951 [00:14<00:16, 57.67it/s]
Loading weights:  51%|#####     | 992/1951 [00:14<00:14, 66.86it/s]
Loading weights:  51%|#####1    | 1001/1951 [00:14<00:14, 63.96it/s]
Loading weights:  52%|#####1    | 1013/1951 [00:15<00:14, 64.28it/s]
Loading weights:  52%|#####2    | 1021/1951 [00:15<00:16, 56.49it/s]
Loading weights:  53%|#####2    | 1028/1951 [00:15<00:18, 49.77it/s]
Loading weights:  53%|#####2    | 1034/1951 [00:15<00:17, 51.13it/s]
Loading weights:  53%|#####3    | 1041/1951 [00:15<00:17, 52.50it/s]
Loading weights:  54%|#####3    | 1047/1951 [00:15<00:19, 45.90it/s]
Loading weights:  54%|#####4    | 1055/1951 [00:16<00:17, 52.01it/s]
Loading weights:  54%|#####4    | 1061/1951 [00:16<00:19, 46.74it/s]
Loading weights:  55%|#####4    | 1069/1951 [00:16<00:16, 53.43it/s]
Loading weights:  55%|#####5    | 1075/1951 [00:16<00:17, 50.19it/s]
Loading weights:  56%|#####5    | 1083/1951 [00:16<00:16, 53.34it/s]
Loading weights:  56%|#####5    | 1089/1951 [00:16<00:18, 45.88it/s]
Loading weights:  56%|#####6    | 1097/1951 [00:16<00:16, 50.76it/s]
Loading weights:  57%|#####6    | 1103/1951 [00:17<00:18, 46.33it/s]
Loading weights:  57%|#####6    | 1111/1951 [00:17<00:16, 49.47it/s]
Loading weights:  57%|#####7    | 1117/1951 [00:17<00:18, 45.23it/s]
Loading weights:  58%|#####7    | 1125/1951 [00:17<00:16, 49.95it/s]
Loading weights:  58%|#####7    | 1131/1951 [00:17<00:17, 45.70it/s]
Loading weights:  58%|#####8    | 1139/1951 [00:17<00:15, 50.97it/s]
Loading weights:  59%|#####8    | 1145/1951 [00:17<00:18, 44.31it/s]
Loading weights:  59%|#####9    | 1153/1951 [00:18<00:16, 47.12it/s]
Loading weights:  59%|#####9    | 1158/1951 [00:18<00:19, 39.84it/s]
Loading weights:  60%|#####9    | 1167/1951 [00:18<00:17, 45.78it/s]
Loading weights:  60%|######    | 1172/1951 [00:18<00:19, 39.82it/s]
Loading weights:  61%|######    | 1181/1951 [00:18<00:16, 47.47it/s]
Loading weights:  61%|######    | 1187/1951 [00:18<00:18, 41.27it/s]
Loading weights:  61%|######1   | 1195/1951 [00:19<00:16, 46.21it/s]
Loading weights:  62%|######1   | 1200/1951 [00:19<00:19, 38.02it/s]
Loading weights:  62%|######1   | 1209/1951 [00:19<00:16, 43.86it/s]
Loading weights:  62%|######2   | 1214/1951 [00:19<00:20, 36.48it/s]
Loading weights:  63%|######2   | 1223/1951 [00:19<00:17, 42.09it/s]
Loading weights:  63%|######2   | 1228/1951 [00:19<00:19, 36.60it/s]
Loading weights:  63%|######3   | 1237/1951 [00:20<00:16, 44.38it/s]
Loading weights:  64%|######3   | 1242/1951 [00:20<00:17, 39.93it/s]
Loading weights:  64%|######4   | 1251/1951 [00:20<00:14, 47.69it/s]
Loading weights:  64%|######4   | 1257/1951 [00:20<00:15, 44.28it/s]
Loading weights:  65%|######4   | 1265/1951 [00:20<00:13, 50.14it/s]
Loading weights:  65%|######5   | 1271/1951 [00:20<00:14, 45.54it/s]
Loading weights:  66%|######5   | 1279/1951 [00:21<00:14, 47.84it/s]
Loading weights:  66%|######5   | 1284/1951 [00:21<00:16, 40.53it/s]
Loading weights:  66%|######6   | 1292/1951 [00:21<00:14, 46.99it/s]
Loading weights:  68%|######8   | 1330/1951 [00:21<00:05, 121.43it/s]
Loading weights:  71%|#######   | 1378/1951 [00:21<00:02, 208.06it/s]
Loading weights:  73%|#######3  | 1425/1951 [00:21<00:01, 268.59it/s]
Loading weights:  75%|#######5  | 1471/1951 [00:21<00:01, 313.37it/s]
Loading weights:  78%|#######8  | 1526/1951 [00:21<00:01, 369.87it/s]
Loading weights:  81%|########  | 1578/1951 [00:21<00:00, 403.80it/s]
Loading weights:  84%|########3 | 1635/1951 [00:22<00:00, 437.75it/s]
Loading weights:  87%|########6 | 1690/1951 [00:22<00:00, 462.98it/s]
Loading weights:  89%|########9 | 1741/1951 [00:22<00:00, 476.08it/s]
Loading weights:  92%|#########1| 1791/1951 [00:22<00:00, 482.60it/s]
Loading weights:  94%|#########4| 1840/1951 [00:22<00:00, 471.82it/s]
Loading weights:  97%|#########7| 1901/1951 [00:22<00:00, 511.38it/s]
Loading weights: 100%|##########| 1951/1951 [00:22<00:00, 85.93it/s] 
[pd] FRESH LoRA r=8 alpha=16 (low-rank grokking substrate)
[pd] no 'prior_weights' column -> forcing --uniform (plain CE weighting)
[pd] data=experiments/v8_nla_local/data/stage3_balanced/av_sft_balanced.parquet rows=1356 uniform=True gamma=1.0 corpus_mean_weight=1.000
[pd] DOMAIN-AWARE negs: avg same-dom=16.0 diff-dom=16.0 per anchor (15 domains)
[pd] contrastive ON: beta=1.0 K=2 temp=1.0 domain_aware=True token_dropout=0.0 (InfoNCE over hard-neg activations)
[pd] INJECT at residual layer 23 (native scale); embed pass-through
[pd] inject_layer=23 inject_mode=raw
[pd] optimizer=adamw8bit lr=5e-05 wd=0.01 steps=1000 ga=1 trainable_params=12.08M
`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.
C:\Users\caleb\deception-nanochat-sae-research\.venv-gemma4\Lib\site-packages\torch\_dynamo\eval_frame.py:1181: UserWarning: torch.utils.checkpoint: the use_reentrant parameter should be passed explicitly. Starting in PyTorch 2.9, calling checkpoint without use_reentrant will raise an exception. use_reentrant=False is recommended, but if you need to preserve the current default behavior, you can pass use_reentrant=True. Refer to docs for more details on the differences between the two variants.
  return fn(*args, **kwargs)
C:\Users\caleb\deception-nanochat-sae-research\experiments\v8_nla_local\train_prior_deviation_av.py:325: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
Consider using tensor.detach() first. (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\torch\csrc\autograd\generated\python_variable_methods.cpp:837.)
  (loss / ga).backward(); tot += float(loss) / ga
[pd] step 1/1000 wloss=5.580 skipped=0
[pd] step 10/1000 wloss=3.317 skipped=0
[pd] step 20/1000 wloss=5.289 skipped=0
[pd] step 30/1000 wloss=3.477 skipped=0
[pd] step 40/1000 wloss=4.236 skipped=0
[pd] step 50/1000 wloss=5.188 skipped=0
[pd] step 60/1000 wloss=3.750 skipped=0
[pd] step 70/1000 wloss=3.908 skipped=0
[pd] step 80/1000 wloss=4.209 skipped=0
[pd] step 90/1000 wloss=5.104 skipped=0
[pd] step 100/1000 wloss=4.671 skipped=0
C:\Users\caleb\deception-nanochat-sae-research\.venv-gemma4\Lib\site-packages\peft\utils\save_and_load.py:372: UserWarning: Could not find a config file in google/gemma-4-E2B - will assume that the vocabulary was not modified.
  warnings.warn(
[pd] saved experiments/v8_nla_local/checkpoints/av_L23_clean\step_000100
[pd] step 110/1000 wloss=4.033 skipped=0
[pd] step 120/1000 wloss=4.942 skipped=0
[pd] step 130/1000 wloss=4.147 skipped=0
[pd] step 140/1000 wloss=3.960 skipped=0
[pd] step 150/1000 wloss=3.524 skipped=0
[pd] step 160/1000 wloss=3.136 skipped=0
[pd] step 170/1000 wloss=5.005 skipped=0
[pd] step 180/1000 wloss=3.400 skipped=0
[pd] step 190/1000 wloss=3.929 skipped=0
[pd] step 200/1000 wloss=4.085 skipped=0
[pd] saved experiments/v8_nla_local/checkpoints/av_L23_clean\step_000200
[pd] step 210/1000 wloss=4.888 skipped=0
[pd] step 220/1000 wloss=3.767 skipped=0
[pd] step 230/1000 wloss=3.771 skipped=0
[pd] step 240/1000 wloss=4.061 skipped=0
[pd] step 250/1000 wloss=3.724 skipped=0
[pd] step 260/1000 wloss=4.040 skipped=0
[pd] step 270/1000 wloss=5.344 skipped=0
[pd] step 280/1000 wloss=3.331 skipped=0
[pd] step 290/1000 wloss=4.078 skipped=0
[pd] step 300/1000 wloss=4.956 skipped=0
[pd] saved experiments/v8_nla_local/checkpoints/av_L23_clean\step_000300
[pd] step 310/1000 wloss=4.509 skipped=0
[pd] step 320/1000 wloss=4.247 skipped=0
[pd] step 330/1000 wloss=4.128 skipped=0
[pd] step 340/1000 wloss=4.066 skipped=0
[pd] step 350/1000 wloss=5.392 skipped=0
[pd] step 360/1000 wloss=3.563 skipped=0
[pd] step 370/1000 wloss=3.632 skipped=0
[pd] step 380/1000 wloss=3.913 skipped=0
[pd] step 390/1000 wloss=3.936 skipped=0
[pd] step 400/1000 wloss=3.896 skipped=0
[pd] saved experiments/v8_nla_local/checkpoints/av_L23_clean\step_000400
[pd] step 410/1000 wloss=3.727 skipped=0
[pd] step 420/1000 wloss=4.693 skipped=0
[pd] step 430/1000 wloss=4.753 skipped=0
[pd] step 440/1000 wloss=3.553 skipped=0
[pd] step 450/1000 wloss=3.824 skipped=0
[pd] step 460/1000 wloss=3.776 skipped=0
[pd] step 470/1000 wloss=3.735 skipped=0