train 737720 lines · val 38578 lines charset extended by 3: ĦƏⱩ charset: 396 characters (min freq 20) dropped 62223 train / 3244 val lines containing unspellable characters /home/danial/.cache/claude-worktrees/bagrat-fund-old-scripts/services/ocr-lab/venv/lib/python3.12/site-packages/torch/cuda/__init__.py:829: UserWarning: Can't initialize NVML warnings.warn("Can't initialize NVML") grew the CTC head 394 → 397 classes (3 new letters) resumed from runs/rec-4script8/init.pt epoch 1 step 200/21109 loss 0.0309 (54s) epoch 1 step 400/21109 loss 0.0286 (106s) epoch 1 step 600/21109 loss 0.0282 (156s) epoch 1 step 800/21109 loss 0.0283 (207s) epoch 1 step 1000/21109 loss 0.0286 (258s) epoch 1 step 1200/21109 loss 0.0292 (308s) epoch 1 step 1400/21109 loss 0.0298 (358s) epoch 1 step 1600/21109 loss 0.0301 (407s) epoch 1 step 1800/21109 loss 0.0308 (458s) epoch 1 step 2000/21109 loss 0.0314 (508s) epoch 1 step 2200/21109 loss 0.0320 (558s) epoch 1 step 2400/21109 loss 0.0323 (608s) epoch 1 step 2600/21109 loss 0.0323 (658s) epoch 1 step 2800/21109 loss 0.0322 (707s) epoch 1 step 3000/21109 loss 0.0327 (756s) epoch 1 step 3200/21109 loss 0.0326 (805s) epoch 1 step 3400/21109 loss 0.0325 (854s) epoch 1 step 3600/21109 loss 0.0327 (903s) epoch 1 step 3800/21109 loss 0.0327 (952s) epoch 1 step 4000/21109 loss 0.0331 (1001s) epoch 1 step 4200/21109 loss 0.0331 (1051s) epoch 1 step 4400/21109 loss 0.0331 (1099s) epoch 1 step 4600/21109 loss 0.0333 (1148s) epoch 1 step 4800/21109 loss 0.0333 (1197s) epoch 1 step 5000/21109 loss 0.0332 (1245s) epoch 1 step 5200/21109 loss 0.0334 (1295s) epoch 1 step 5400/21109 loss 0.0337 (1343s) epoch 1 step 5600/21109 loss 0.0336 (1391s) epoch 1 step 5800/21109 loss 0.0339 (1440s) epoch 1 step 6000/21109 loss 0.0339 (1488s) epoch 1 step 6200/21109 loss 0.0339 (1537s) epoch 1 step 6400/21109 loss 0.0338 (1586s) epoch 1 step 6600/21109 loss 0.0338 (1636s) epoch 1 step 6800/21109 loss 0.0340 (1685s) epoch 1 step 7000/21109 loss 0.0338 (1733s) epoch 1 step 7200/21109 loss 0.0338 (1782s) epoch 1 step 7400/21109 loss 0.0339 (1831s) epoch 1 step 7600/21109 loss 0.0340 (1880s) epoch 1 step 7800/21109 loss 0.0341 (1929s) epoch 1 step 8000/21109 loss 0.0342 (1977s) epoch 1 step 8200/21109 loss 0.0339 (2026s) epoch 1 step 8400/21109 loss 0.0338 (2074s) epoch 1 step 8600/21109 loss 0.0340 (2123s) epoch 1 step 8800/21109 loss 0.0339 (2171s) epoch 1 step 9000/21109 loss 0.0338 (2220s) epoch 1 step 9200/21109 loss 0.0338 (2268s) epoch 1 step 9400/21109 loss 0.0338 (2317s) epoch 1 step 9600/21109 loss 0.0338 (2365s) epoch 1 step 9800/21109 loss 0.0337 (2414s) epoch 1 step 10000/21109 loss 0.0336 (2462s) epoch 1 step 10200/21109 loss 0.0335 (2510s) epoch 1 step 10400/21109 loss 0.0334 (2558s) epoch 1 step 10600/21109 loss 0.0333 (2606s) epoch 1 step 10800/21109 loss 0.0333 (2655s) epoch 1 step 11000/21109 loss 0.0333 (2703s) epoch 1 step 11200/21109 loss 0.0333 (2752s) epoch 1 step 11400/21109 loss 0.0333 (2800s) epoch 1 step 11600/21109 loss 0.0332 (2849s) epoch 1 step 11800/21109 loss 0.0333 (2897s) epoch 1 step 12000/21109 loss 0.0332 (2945s) epoch 1 step 12200/21109 loss 0.0332 (2994s) epoch 1 step 12400/21109 loss 0.0332 (3042s) epoch 1 step 12600/21109 loss 0.0331 (3090s) epoch 1 step 12800/21109 loss 0.0331 (3139s) epoch 1 step 13000/21109 loss 0.0331 (3187s) epoch 1 step 13200/21109 loss 0.0330 (3236s) epoch 1 step 13400/21109 loss 0.0331 (3284s) epoch 1 step 13600/21109 loss 0.0330 (3332s) epoch 1 step 13800/21109 loss 0.0329 (3381s) epoch 1 step 14000/21109 loss 0.0328 (3429s) epoch 1 step 14200/21109 loss 0.0327 (3478s) epoch 1 step 14400/21109 loss 0.0328 (3526s) epoch 1 step 14600/21109 loss 0.0327 (3575s) epoch 1 step 14800/21109 loss 0.0326 (3623s) epoch 1 step 15000/21109 loss 0.0328 (3672s) epoch 1 step 15200/21109 loss 0.0328 (3720s) epoch 1 step 15400/21109 loss 0.0328 (3768s) epoch 1 step 15600/21109 loss 0.0328 (3817s) epoch 1 step 15800/21109 loss 0.0328 (3865s) epoch 1 step 16000/21109 loss 0.0328 (3914s) epoch 1 step 16200/21109 loss 0.0327 (3962s) epoch 1 step 16400/21109 loss 0.0327 (4011s) epoch 1 step 16600/21109 loss 0.0326 (4059s) epoch 1 step 16800/21109 loss 0.0327 (4107s) epoch 1 step 17000/21109 loss 0.0326 (4155s) epoch 1 step 17200/21109 loss 0.0325 (4203s) epoch 1 step 17400/21109 loss 0.0325 (4252s) epoch 1 step 17600/21109 loss 0.0325 (4300s) epoch 1 step 17800/21109 loss 0.0325 (4349s) epoch 1 step 18000/21109 loss 0.0324 (4397s) epoch 1 step 18200/21109 loss 0.0324 (4445s) epoch 1 step 18400/21109 loss 0.0324 (4493s) epoch 1 step 18600/21109 loss 0.0324 (4542s) epoch 1 step 18800/21109 loss 0.0324 (4590s) epoch 1 step 19000/21109 loss 0.0323 (4638s) epoch 1 step 19200/21109 loss 0.0323 (4687s) epoch 1 step 19400/21109 loss 0.0322 (4735s) epoch 1 step 19600/21109 loss 0.0322 (4783s) epoch 1 step 19800/21109 loss 0.0321 (4831s) epoch 1 step 20000/21109 loss 0.0321 (4879s) epoch 1 step 20200/21109 loss 0.0320 (4928s) epoch 1 step 20400/21109 loss 0.0320 (4977s) epoch 1 step 20600/21109 loss 0.0320 (5026s) epoch 1 step 20800/21109 loss 0.0320 (5074s) epoch 1 step 21000/21109 loss 0.0320 (5123s) epoch 1/1 loss 0.0320 CER verbatim 0.0105 normalized 0.0103 (5229s) ↑ best CER 0.0103 done. best normalized CER 0.0103 (gate: ≤ 0.10)