{ "run_name": "gpu2_b2048_edge_finetune", "training_seconds": 561.47, "best_step": 500, "best_eval_exact_accuracy": 0.993, "final_exact_accuracy": 0.9929, "final_exact_correct": 19858, "final_exact_total": 20000, "final_exact_by_operation": { "+": { "correct": 4022, "total": 4022, "accuracy": 1.0 }, "-": { "correct": 3965, "total": 3965, "accuracy": 1.0 }, "*": { "correct": 4084, "total": 4084, "accuracy": 1.0 }, "//": { "correct": 4007, "total": 4028, "accuracy": 0.9947864945382324 }, "%": { "correct": 3780, "total": 3901, "accuracy": 0.968982312227634 } }, "model_path": "runs/gpu2_b2048_edge_finetune/arithmetic_transformer_params.msgpack", "best_model_path": "runs/gpu2_b2048_edge_finetune/best_params.msgpack", "runtime": { "timestamp": "2026-05-22 20:47:17 KST", "python": "3.13.13", "platform": "Linux-5.4.0-216-generic-x86_64-with-glibc2.31", "jax": "0.10.1", "backend": "gpu", "devices": [ "cuda:0" ], "cuda_visible_devices": "2", "nvidia_smi": "Fri May 22 20:47:17 2026 \n+-----------------------------------------------------------------------------------------+\n| NVIDIA-SMI 575.57.08 Driver Version: 575.57.08 CUDA Version: 12.9 |\n|-----------------------------------------+------------------------+----------------------+\n| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |\n| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |\n| | | MIG M. |\n|=========================================+========================+======================|\n| 0 NVIDIA GeForce RTX 3090 Off | 00000000:1B:00.0 Off | N/A |\n| 54% 32C P8 25W / 350W | 22932MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 1 NVIDIA GeForce RTX 3090 Off | 00000000:1C:00.0 Off | N/A |\n| 43% 30C P8 25W / 350W | 22430MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 2 NVIDIA GeForce RTX 3090 Off | 00000000:1D:00.0 Off | N/A |\n| 54% 37C P2 38W / 350W | 272MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 3 NVIDIA GeForce RTX 3090 Off | 00000000:1E:00.0 Off | N/A |\n| 53% 33C P8 26W / 350W | 1MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 4 NVIDIA GeForce RTX 3090 Off | 00000000:3D:00.0 Off | N/A |\n| 51% 30C P8 27W / 350W | 1MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 5 NVIDIA GeForce RTX 3090 Off | 00000000:3F:00.0 Off | N/A |\n| 53% 32C P8 23W / 350W | 1MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 6 NVIDIA GeForce RTX 3090 Off | 00000000:40:00.0 Off | N/A |\n| 53% 31C P8 26W / 350W | 1MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n| 7 NVIDIA GeForce RTX 3090 Off | 00000000:41:00.0 Off | N/A |\n| 52% 31C P8 24W / 350W | 1MiB / 24576MiB | 0% Default |\n| | | N/A |\n+-----------------------------------------+------------------------+----------------------+\n \n+-----------------------------------------------------------------------------------------+\n| Processes: |\n| GPU GI CI PID Type Process name GPU Memory |\n| ID ID Usage |\n|=========================================================================================|\n| 0 N/A N/A 16627 C VLLM::EngineCore 22922MiB |\n| 1 N/A N/A 16558 C VLLM::EngineCore 22420MiB |\n| 2 N/A N/A 1956597 C .miniforge/bin/python 262MiB |\n+-----------------------------------------------------------------------------------------+" }, "dataset": { "operations": [ "+", "-", "*", "//", "%" ], "total_examples": 4998000, "unique_valid_examples": 4998000, "train_examples": 4898000, "validation_examples": 100000, "train_examples_by_operation": { "+": 980141, "-": 980000, "*": 979815, "//": 978986, "%": 979058 }, "validation_examples_by_operation": { "+": 19859, "-": 20000, "*": 20185, "//": 20014, "%": 19942 }, "sequence_length": 28, "prompt_length": 9, "answer_length": 19, "vocab_size": 17, "vocab": "0123456789 +-*%/=", "dataset_build_seconds": 57.91, "dataset_memory_mb": 559.78, "division_modulo_zero_policy": "b=0 is omitted for // and %" }, "model": { "vocab_size": 17, "max_seq_len": 28, "d_model": 384, "n_heads": 6, "n_layers": 6, "mlp_dim": 1536, "parameter_count": 10650624, "parameter_count_millions": 10.651 }, "hyperparameters": { "run_name": "gpu2_b2048_edge_finetune", "output_dir": "runs", "batch_size": 2048, "eval_batch_size": 256, "max_steps": 1500, "min_steps": 500, "eval_every": 250, "eval_exact_n": 1000, "final_exact_n": 20000, "target_exact_accuracy": 0.995, "learning_rate": 0.0002, "end_lr_ratio": 0.1, "weight_decay": 0.0001, "warmup_steps": 50, "grad_clip": 1.0, "multiplication_focus_steps": 2500, "multiplication_focus_prob": 0.7, "op_probs": [ 0.05, 0.1, 0.2, 0.3, 0.35 ], "seed": 1234, "init_seed": 0, "d_model": 384, "n_heads": 6, "n_layers": 6, "mlp_dim": 1536, "val_size": 100000, "save_every_eval": false, "require_gpu": true, "init_params": "runs/gpu3_b2048_divmod5000/arithmetic_transformer_params.msgpack", "edge_case_prob": 0.25 }, "sample_predictions": [ { "problem": "0+0", "model_text": " 0+ 0=+000000000000000000", "prediction": "0", "expected": "0", "correct": true }, { "problem": "7+35", "model_text": " 7+ 35=+200400000000000000", "prediction": "42", "expected": "42", "correct": true }, { "problem": "123-45", "model_text": "123- 45=+800700000000000000", "prediction": "78", "expected": "78", "correct": true }, { "problem": "45-123", "model_text": " 45- 123=-800700000000000000", "prediction": "-78", "expected": "-78", "correct": true }, { "problem": "12*89", "model_text": " 12* 89=+801602001100000000", "prediction": "1068", "expected": "1068", "correct": true }, { "problem": "123*45", "model_text": "123* 45=+501302501500000000", "prediction": "5535", "expected": "5535", "correct": true }, { "problem": "5//4", "model_text": " 5// 4=+100000000000000000", "prediction": "1", "expected": "1", "correct": true }, { "problem": "999//1", "model_text": "999// 1=+900900900000000000", "prediction": "999", "expected": "999", "correct": true }, { "problem": "5%4", "model_text": " 5% 4=+100000000000000000", "prediction": "1", "expected": "1", "correct": true }, { "problem": "998%999", "model_text": "998% 999=+800900900000000000", "prediction": "998", "expected": "998", "correct": true }, { "problem": "501*499", "model_text": "501* 499=+900900904904402200", "prediction": "249999", "expected": "249999", "correct": true }, { "problem": "999*999", "model_text": "999* 999=+108017026818909900", "prediction": "998001", "expected": "998001", "correct": true } ], "history": [ { "event": "eval", "step": 1, "elapsed_seconds": 277.63, "train_loss": 0.002636338584125042, "train_token_accuracy": 0.9991262555122375, "val_loss": 0.001127233263105154, "val_token_accuracy": 0.9994603395462036, "exact_accuracy": 0.99, "exact_correct": 990, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 194, "total": 194, "accuracy": 1.0 }, "-": { "correct": 198, "total": 198, "accuracy": 1.0 }, "*": { "correct": 205, "total": 205, "accuracy": 1.0 }, "//": { "correct": 198, "total": 200, "accuracy": 0.99 }, "%": { "correct": 195, "total": 203, "accuracy": 0.9605911330049262 } }, "learning_rate": 4.0000013541430235e-06, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 250, "elapsed_seconds": 324.4, "train_loss": 0.004352588206529617, "train_token_accuracy": 0.9986379742622375, "val_loss": 0.002333001233637333, "val_token_accuracy": 0.9990748167037964, "exact_accuracy": 0.983, "exact_correct": 983, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 174, "total": 174, "accuracy": 1.0 }, "-": { "correct": 197, "total": 197, "accuracy": 1.0 }, "*": { "correct": 181, "total": 183, "accuracy": 0.9890710382513661 }, "//": { "correct": 222, "total": 226, "accuracy": 0.9823008849557522 }, "%": { "correct": 209, "total": 220, "accuracy": 0.95 } }, "learning_rate": 0.00019168177095707506, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 500, "elapsed_seconds": 371.57, "train_loss": 0.0026333255227655172, "train_token_accuracy": 0.9988692402839661, "val_loss": 0.0016257410170510411, "val_token_accuracy": 0.9993318319320679, "exact_accuracy": 0.993, "exact_correct": 993, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 203, "total": 203, "accuracy": 1.0 }, "-": { "correct": 210, "total": 210, "accuracy": 1.0 }, "*": { "correct": 205, "total": 205, "accuracy": 1.0 }, "//": { "correct": 188, "total": 191, "accuracy": 0.9842931937172775 }, "%": { "correct": 187, "total": 191, "accuracy": 0.9790575916230366 } }, "learning_rate": 0.00016050682461354882, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 750, "elapsed_seconds": 419.21, "train_loss": 0.0022381972521543503, "train_token_accuracy": 0.9991776347160339, "val_loss": 0.0018789999885484576, "val_token_accuracy": 0.9992290139198303, "exact_accuracy": 0.99, "exact_correct": 990, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 202, "total": 202, "accuracy": 1.0 }, "-": { "correct": 198, "total": 198, "accuracy": 1.0 }, "*": { "correct": 193, "total": 193, "accuracy": 1.0 }, "//": { "correct": 206, "total": 209, "accuracy": 0.9856459330143541 }, "%": { "correct": 191, "total": 198, "accuracy": 0.9646464646464646 } }, "learning_rate": 0.00011487249139463529, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 1000, "elapsed_seconds": 466.53, "train_loss": 0.0014942348934710026, "train_token_accuracy": 0.9993061423301697, "val_loss": 0.0012976628495380282, "val_token_accuracy": 0.9994603395462036, "exact_accuracy": 0.993, "exact_correct": 993, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 214, "total": 214, "accuracy": 1.0 }, "-": { "correct": 195, "total": 195, "accuracy": 1.0 }, "*": { "correct": 176, "total": 176, "accuracy": 1.0 }, "//": { "correct": 209, "total": 210, "accuracy": 0.9952380952380953 }, "%": { "correct": 199, "total": 205, "accuracy": 0.9707317073170731 } }, "learning_rate": 6.784322613384575e-05, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 1250, "elapsed_seconds": 513.89, "train_loss": 0.0016579929506406188, "train_token_accuracy": 0.9992804527282715, "val_loss": 0.0017736267764121294, "val_token_accuracy": 0.9992547631263733, "exact_accuracy": 0.989, "exact_correct": 989, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 199, "total": 199, "accuracy": 1.0 }, "-": { "correct": 210, "total": 210, "accuracy": 1.0 }, "*": { "correct": 194, "total": 194, "accuracy": 1.0 }, "//": { "correct": 197, "total": 198, "accuracy": 0.9949494949494949 }, "%": { "correct": 189, "total": 199, "accuracy": 0.949748743718593 } }, "learning_rate": 3.288284642621875e-05, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } }, { "event": "eval", "step": 1500, "elapsed_seconds": 561.47, "train_loss": 0.0014991366770118475, "train_token_accuracy": 0.9993318319320679, "val_loss": 0.0011124522425234318, "val_token_accuracy": 0.9995630979537964, "exact_accuracy": 0.993, "exact_correct": 993, "exact_total": 1000, "exact_by_operation": { "+": { "correct": 191, "total": 191, "accuracy": 1.0 }, "-": { "correct": 208, "total": 209, "accuracy": 0.9952153110047847 }, "*": { "correct": 192, "total": 192, "accuracy": 1.0 }, "//": { "correct": 222, "total": 224, "accuracy": 0.9910714285714286 }, "%": { "correct": 180, "total": 184, "accuracy": 0.9782608695652174 } }, "learning_rate": 1.9999999494757503e-05, "op_sampling": { "+": 0.05, "-": 0.1, "*": 0.2, "//": 0.3, "%": 0.35 } } ] }