2025-09-13 23:07:23,868 - INFO - Arguments: Namespace(output_dir='./marvis_fft_results/task_32/split_2/model', seed=44, resume=None, resume_optimizer=False, model_id='Qwen/Qwen2.5-3B-Instruct', embedding_size=1000, unfreeze_last_k_layers=1, use_vector_quantization=False, vq_num_embeddings=256, vq_commitment_cost=1.0, vq_decay=0.99, vq_debug=False, vq_warmup_steps=100, total_steps=2000, batch_size=8, max_train_samples=None, max_test_samples=None, learning_rate=0.0001, mixup_alpha=0.0, gradient_accumulation_steps=1, save_best_model=True, save_steps=500, save_total_limit=None, lr_scheduler='cosine', lr_warmup_steps=0, lr_warmup_ratio=0.0, lr_eta_min=1e-06, lr_gamma=0.1, lr_power=1.0, lr_num_cycles=0.5, lr_cycle_length=None, vq_learning_rate=None, vq_lr_scheduler=None, class_token_learning_rate=None, class_token_lr_scheduler=None, llm_learning_rate=None, llm_lr_scheduler=None, temperature_initial=1.0, temperature_final=0.1, temperature_anneal_steps=None, gradient_penalty_weight=0.0, gradient_penalty_threshold=10.0, embedding_cache_dir='./data', force_recompute_embeddings=False, allow_small_datasets=True, num_few_shot_examples=100, permute_examples=False, permute_labels=False, permute_labels_every_k_steps=None, no_permute_last_k=None, variable_few_shot=False, few_shot_min=10, few_shot_max=None, min_freq_weight=0.05, min_freq_target=0.05, early_stopping_patience=30, use_wandb=True, resume_wandb=False, wandb_project='marvis-openml-cc18-v20250913', wandb_entity='nyu-dice-lab', wandb_name='train_task32_split2', dataset_name='32', num_epochs=3, openml_task_id=None, early_stopping_threshold=0.4, bypass_eval=True, evaluate_with_explanations=False, explanation_type='standard', max_explanation_tokens=50, save_explanations=False)
2025-09-13 23:07:24,755 - INFO - CUDA available with 1 GPU(s)
2025-09-13 23:07:24,759 - INFO - Started GPU monitoring with 30.0s interval
2025-09-13 23:07:24,762 - INFO - Initialized Weights & Biases run: train_task32_split2 (ID: gwtppgag)
2025-09-13 23:07:24,762 - INFO - Loading dataset
2025-09-13 23:07:24,767 - INFO - Saved 0 failed dataset IDs to managed cache
2025-09-13 23:07:24,767 - INFO - Cleared the failed dataset cache
2025-09-13 23:07:24,767 - INFO - Cleared _FAILED_DATASET_CACHE to ensure dataset loading is attempted
2025-09-13 23:07:24,868 - INFO - Loading dataset with ID 32 from OpenML
2025-09-13 23:07:24,872 - INFO - Successfully loaded dataset: pendigits (ID: 32)
2025-09-13 23:07:24,898 - INFO - Dataset has default target attribute: 'class'
2025-09-13 23:07:24,898 - INFO - Converting categorical target to numeric codes
2025-09-13 23:07:24,900 - INFO - Cleaning NaN and inf values from dataset
2025-09-13 23:07:24,901 - INFO - Final dataset size after cleaning: 10992 samples
2025-09-13 23:07:24,901 - INFO - Dataset info: 10992 samples, 16 features (0 categorical, 16 numerical), 10 classes
2025-09-13 23:07:24,901 - INFO - Loaded dataset: pendigits
2025-09-13 23:07:24,902 - INFO - Identifiers — Dataset ID: 32, Task ID: unknown
2025-09-13 23:07:24,904 - INFO - Dataset shapes - Train: (7694, 16), Val: (1649, 16), Test: (1649, 16)
2025-09-13 23:07:24,904 - INFO - Getting TabPFN embeddings with size 1000
2025-09-13 23:07:24,905 - INFO - Using embedding cache directory: ./data
2025-09-13 23:07:25,131 - INFO - Fitting TabPFN and extracting embeddings
2025-09-13 23:07:25,131 - INFO - Original train set class distribution: {0: 805, 1: 778, 2: 787, 3: 758, 4: 807, 5: 741, 6: 727, 7: 801, 8: 756, 9: 734}
2025-09-13 23:07:25,135 - INFO - Sampled 3000 training samples from 7694 total samples
2025-09-13 23:07:25,135 - INFO - Sampled train set class distribution: {0: 314, 1: 303, 2: 307, 3: 296, 4: 315, 5: 289, 6: 283, 7: 312, 8: 295, 9: 286}
2025-09-13 23:07:25,135 - INFO - Initializing TabPFN on device: cuda
2025-09-13 23:07:25,135 - INFO - Starting TabPFN.fit on 3000 samples, device=cuda
2025-09-13 23:07:25,537 - INFO - TabPFN.fit completed in 0.40s
2025-09-13 23:07:25,538 - INFO - Computing TabPFN train embeddings...
2025-09-13 23:07:26,678 - INFO - TabPFN train embeddings computed in 1.14s
2025-09-13 23:07:26,679 - INFO - Computing TabPFN validation and test embeddings in chunks...
2025-09-13 23:07:27,735 - INFO - Raw embedding shapes - Train: (8, 3000, 192), Val: (8, 1649, 192), Test: (8, 1649, 192)
2025-09-13 23:07:27,735 - INFO - Averaging embeddings across ensemble members
2025-09-13 23:07:27,758 - INFO - Processed embedding shapes - Train: (3000, 192), Val: (1649, 192), Test: (1649, 192)
2025-09-13 23:07:27,810 - INFO - RobustScaler statistics - Center (median): [ 0.2612  -0.0519  -0.3733  -0.1802  -0.10504]... Scale (IQR): [0.32797241 0.23938751 0.37640381 0.31143188 0.24072838]...
2025-09-13 23:07:27,810 - INFO - Resizing embeddings from 192 to 1000
2025-09-13 23:07:27,821 - INFO - Resizing embeddings from 192 to 1000
2025-09-13 23:07:27,830 - INFO - Saving TabPFN embeddings to managed cache
2025-09-13 23:07:27,900 - INFO - Successfully saved embeddings to managed cache
2025-09-13 23:07:27,900 - INFO - Final embedding shapes - Train: (3000, 1000), Val: None, Test: (1649, 1000)
2025-09-13 23:07:27,902 - INFO - Preparing model Qwen/Qwen2.5-3B-Instruct with regular prefix embedding
2025-09-13 23:07:27,902 - INFO - Loading Qwen/Qwen2.5-3B-Instruct model and tokenizer
2025-09-13 23:07:36,736 - INFO - Added special tokens: PREFIX_START_ID=151665, PREFIX_END_ID=151666
2025-09-13 23:07:36,736 - INFO - Class token IDs: [151667, 151668, 151669, 151670, 151671, 151672, 151673, 151674, 151675, 151676]
2025-09-13 23:07:36,744 - INFO - Creating LLM dataset
2025-09-13 23:07:36,744 - INFO - Creating datasets for LLM training and evaluation
2025-09-13 23:07:36,746 - INFO - Encoded 10 unique classes
2025-09-13 23:07:36,746 - INFO - TabPFN embedding shapes - Train: (3000, 1000), Val: (1649, 1000), Test: (1649, 1000)
2025-09-13 23:07:36,747 - INFO - Prepared prefix data - Embeddings: (100, 1000), Class labels: (100,)
2025-09-13 23:07:36,755 - INFO - Saved prefix data to ./marvis_fft_results/task_32/split_2/model/prefix_data.npz
2025-09-13 23:07:36,846 - INFO - Created datasets - Train: 1319, Eval: 330, Test: 1649
2025-09-13 23:07:37,129 - INFO - Training model for 2000 total steps
2025-09-13 23:07:37,130 - INFO - Setting up custom training loop
2025-09-13 23:07:37,134 - INFO - Loaded prefix data from ./marvis_fft_results/task_32/split_2/model/prefix_data.npz
2025-09-13 23:07:37,134 - INFO - Embeddings shape: torch.Size([100, 1000]), Class labels shape: torch.Size([100])
2025-09-13 23:07:37,135 - INFO - Example permutation: DISABLED - Example order will remain fixed
2025-09-13 23:07:37,135 - INFO - Label permutation: DISABLED - Class-to-label mapping will remain fixed
2025-09-13 23:07:37,135 - INFO - Variable few-shot: DISABLED - Using fixed 100 examples
2025-09-13 23:07:37,425 - INFO - Dataset has 10 classes
2025-09-13 23:07:37,427 - INFO - Training for 2000 total steps with learning rate schedule:
2025-09-13 23:07:37,427 - INFO -   - First 800 steps (40.0%): Fixed rate 0.0001
2025-09-13 23:07:37,427 - INFO -   - Remaining steps: Exponential decay from 0.0001 to 1.6000000000000003e-05
2025-09-13 23:07:37,427 - INFO -   - Temperature annealing: 1.0 -> 0.1 over None steps
2025-09-13 23:07:37,427 - INFO -   - Gradient penalty: weight=0.0, threshold=10.0
2025-09-13 23:07:37,427 - INFO -   - Layer freezing: Unfreeze last 1 layers
2025-09-13 23:07:37,427 - INFO -   - Mixup augmentation with alpha=0.0
2025-09-13 23:07:37,427 - INFO -   - Minimum frequency regularization with weight=0.05, target=0.05
2025-09-13 23:07:37,427 - INFO -   - Saving best model by training loss
2025-09-13 23:07:37,429 - INFO - Unfroze last 1 layers (layers 35-35)
2025-09-13 23:07:37,429 - INFO - Also unfroze lm_head (output layer)
2025-09-13 23:07:37,430 - INFO - Layer freezing complete: 389,761,536 / 3,087,458,304 parameters are trainable (12.6%)
2025-09-13 23:07:37,431 - INFO - LLM parameter group: 15 parameters, lr=0.0001
2025-09-13 23:07:37,431 - INFO - Training on device: cuda:0
2025-09-13 23:07:37,432 - INFO - Training mode: full model in train mode
2025-09-13 23:07:37,433 - INFO - Moved embedding_projector to cuda:0
2025-09-13 23:07:37,502 - INFO - Starting training
2025-09-13 23:07:37,790 - INFO - Model supports 10 classes, dataset has 10 unique classes
2025-09-13 23:07:37,790 - INFO - Unique labels in dataset: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
2025-09-13 23:07:37,790 - INFO - Setting up training on device: cuda:0
2025-09-13 23:07:37,790 - INFO - CUDA available with 1 devices
2025-09-13 23:07:37,790 - INFO - CUDA:0 - NVIDIA H100 80GB HBM3
2025-09-13 23:07:37,790 - INFO - Setting num_train_epochs to 13 based on max_steps 2000
2025-09-13 23:09:00,761 - INFO - Saving new best model to ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:09:33,271 - INFO - Saving new best model to ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:10:53,998 - INFO - Saving new best model to ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:11:25,148 - INFO - Saving new best model to ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:11:57,115 - INFO - Saving new best model to ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:13:02,094 - INFO - Early stopping after 30 steps without improvement
2025-09-13 23:13:02,096 - INFO - Best model was at step 204 with average loss 0.0731
2025-09-13 23:13:02,096 - INFO - Best model saved at: ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:13:10,684 - INFO - Loading model with permuted class token mapping
2025-09-13 23:13:10,685 - INFO - Using final epoch class token IDs: [151667, 151668, 151669, 151670, 151671, 151672, 151673, 151674, 151675, 151676]
2025-09-13 23:13:10,716 - INFO - Loaded best model from ./marvis_fft_results/task_32/split_2/model/best_model
2025-09-13 23:13:10,721 - INFO - Training completed in 333.59 seconds
2025-09-13 23:13:10,721 - INFO - Saving final model checkpoint
2025-09-13 23:13:10,724 - INFO - Saving final model to ./marvis_fft_results/task_32/split_2/model/final_model
2025-09-13 23:13:10,724 - INFO - Saving final model with current (possibly permuted) class token mapping
2025-09-13 23:13:21,207 - INFO - Final model saved successfully to ./marvis_fft_results/task_32/split_2/model/final_model
2025-09-13 23:13:21,208 - INFO - Final model saved to ./marvis_fft_results/task_32/split_2/model/final_model
2025-09-13 23:13:21,208 - INFO - Skipping evaluation as requested with --bypass_eval