{ "run": "main_expand_copy_s750", "arm": "expand_copy", "args": { "run_name": "main_expand_copy_s750", "arm": "expand_copy", "model_id": "Qwen/Qwen3-4B", "seed": 123, "modulus": 97, "stage_max_steps": "8,16,24", "stage_min_lengths": "1,1,8", "stage_train_max_lengths": "8,16,24", "stage_steps": "300,150,300", "train_examples": 512, "train_template_mode": "mixed", "eval_lengths": "8,16,24", "eval_examples": 64, "paired_eval_pairs": 32, "batch_size": 8, "grad_accum": 1, "eval_batch_size": 8, "max_length": 2048, "head_width": 512, "compiler_layers": 1, "compiler_heads": 4, "compiler_dropout": 0.05, "expansion_noise": 0.005, "expansion_mode": "copy_last", "trace_loss_weight": 1.0, "executor_loss_weight": 1.0, "state_loss_weight": 1.0, "init_trace_loss_weight": 4.0, "op_trace_loss_weight": 1.0, "arg_trace_loss_weight": 4.0, "direct_head_weight": 0.0, "lr": 0.0002, "weight_decay": 0.0, "optimizer": "paged_adamw_8bit", "max_grad_norm": 1.0, "torch_dtype": "bf16", "load_in_4bit": true, "device_map": "auto", "use_lora": true, "gradient_checkpointing": true, "lora_r": 8, "lora_alpha": 16, "lora_dropout": 0.05, "lora_target_modules": "all-linear", "log_interval": 50, "save_checkpoints": true, "checkpoint_dir": "", "selection_split": "paired_L24", "selection_metric": "executor_pair_both_correct" }, "metadata": { "python": "3.12.3", "platform": "Linux-6.8.0-62-generic-x86_64-with-glibc2.39", "torch": "2.8.0+cu128", "cuda_available": true, "transformers_loader": "AutoModelForCausalLM", "peft_installed": true, "use_lora": true, "lora_r": 8, "model_id": "Qwen/Qwen3-4B", "gpu_name": "NVIDIA RTX 6000 Ada Generation", "gpu_vram_gb": 47.383 }, "stages": [ { "name": "stage1_max8", "max_steps": 8, "train_min_len": 1, "train_max_len": 8, "steps": 300 }, { "name": "stage2_max16", "max_steps": 16, "train_min_len": 1, "train_max_len": 16, "steps": 150 }, { "name": "stage3_max24", "max_steps": 24, "train_min_len": 8, "train_max_len": 24, "steps": 300 } ], "train_log": [ { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 1, "global_step": 1, "compiler_max_steps": 8, "elapsed_sec": 2.623, "expansion_event": "init", "init_loss": 4.664542198181152, "op_loss": 1.146849274635315, "arg_loss": 4.618834972381592, "state_loss": 4.577921390533447, "state_train_accuracy": 0.0, "executor_loss": 4.577940940856934, "loss": 47.436222076416016 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 50, "global_step": 50, "compiler_max_steps": 8, "elapsed_sec": 55.544, "expansion_event": "init", "init_loss": 4.586603164672852, "op_loss": 0.8490704298019409, "arg_loss": 3.6816141605377197, "state_loss": 4.570467472076416, "state_train_accuracy": 0.0, "executor_loss": 4.578827857971191, "loss": 43.07123565673828 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 100, "global_step": 100, "compiler_max_steps": 8, "elapsed_sec": 110.668, "expansion_event": "init", "init_loss": 2.1068058013916016, "op_loss": 0.007378753740340471, "arg_loss": 0.9835323691368103, "state_loss": 3.1181857585906982, "state_train_accuracy": 0.6428571939468384, "executor_loss": 3.4799280166625977, "loss": 18.96684455871582 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 150, "global_step": 150, "compiler_max_steps": 8, "elapsed_sec": 166.638, "expansion_event": "init", "init_loss": 0.3542819023132324, "op_loss": 0.0007297900738194585, "arg_loss": 0.014904552139341831, "state_loss": 0.31338247656822205, "state_train_accuracy": 1.0, "executor_loss": 0.44162312150001526, "loss": 2.2324812412261963 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 200, "global_step": 200, "compiler_max_steps": 8, "elapsed_sec": 222.626, "expansion_event": "init", "init_loss": 0.01224108599126339, "op_loss": 1.9703367797774263e-05, "arg_loss": 0.001109678647480905, "state_loss": 0.017285553738474846, "state_train_accuracy": 1.0, "executor_loss": 0.01915685459971428, "loss": 0.08986517041921616 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 250, "global_step": 250, "compiler_max_steps": 8, "elapsed_sec": 278.638, "expansion_event": "init", "init_loss": 0.3479628562927246, "op_loss": 4.007452389487298e-06, "arg_loss": 0.0002171544765587896, "state_loss": 0.4736778140068054, "state_train_accuracy": 0.8297871947288513, "executor_loss": 0.34923216700553894, "loss": 2.2156341075897217 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage1_max8", "stage_idx": 1, "local_step": 300, "global_step": 300, "compiler_max_steps": 8, "elapsed_sec": 334.937, "expansion_event": "init", "init_loss": 0.0012989494716748595, "op_loss": 1.9896576759492746e-06, "arg_loss": 0.00010719573765527457, "state_loss": 0.0016946335090324283, "state_train_accuracy": 1.0, "executor_loss": 0.0018720910884439945, "loss": 0.009193294681608677 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage2_max16", "stage_idx": 2, "local_step": 1, "global_step": 301, "compiler_max_steps": 16, "elapsed_sec": 355.402, "expansion_event": "expand_8_to_16_copy_last", "init_loss": 0.0007056244648993015, "op_loss": 0.26750412583351135, "arg_loss": 0.04580557718873024, "state_loss": 0.24515192210674286, "state_train_accuracy": 0.9767441749572754, "executor_loss": 1.3121273517608643, "loss": 2.0108282566070557 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage2_max16", "stage_idx": 2, "local_step": 50, "global_step": 350, "compiler_max_steps": 16, "elapsed_sec": 463.368, "expansion_event": "expand_8_to_16_copy_last", "init_loss": 0.0010084665846079588, "op_loss": 2.221352951892186e-05, "arg_loss": 0.00018391771300230175, "state_loss": 0.002133946865797043, "state_train_accuracy": 0.9999999403953552, "executor_loss": 0.002423361409455538, "loss": 0.00934905931353569 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage2_max16", "stage_idx": 2, "local_step": 100, "global_step": 400, "compiler_max_steps": 16, "elapsed_sec": 573.422, "expansion_event": "expand_8_to_16_copy_last", "init_loss": 0.00020769130787812173, "op_loss": 3.344016022310825e-06, "arg_loss": 8.222608448704705e-05, "state_loss": 0.0006558297900483012, "state_train_accuracy": 1.0, "executor_loss": 0.0008278421591967344, "loss": 0.00264668557792902 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage2_max16", "stage_idx": 2, "local_step": 150, "global_step": 450, "compiler_max_steps": 16, "elapsed_sec": 683.651, "expansion_event": "expand_8_to_16_copy_last", "init_loss": 9.540620521875098e-05, "op_loss": 5.8092286963074e-06, "arg_loss": 3.311126056360081e-05, "state_loss": 0.0002799993962980807, "state_train_accuracy": 1.0, "executor_loss": 0.000405889906687662, "loss": 0.0012057684361934662 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 1, "global_step": 451, "compiler_max_steps": 24, "elapsed_sec": 748.054, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 0.00013861153274774551, "op_loss": 0.17999175190925598, "arg_loss": 0.3630619943141937, "state_loss": 0.2835788428783417, "state_train_accuracy": 0.9619048237800598, "executor_loss": 1.4493041038513184, "loss": 3.3656771183013916 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 50, "global_step": 500, "compiler_max_steps": 24, "elapsed_sec": 918.477, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 0.00016083098307717592, "op_loss": 2.3961026727192802e-06, "arg_loss": 8.958354737842456e-05, "state_loss": 0.0009987991070374846, "state_train_accuracy": 1.0, "executor_loss": 0.0015395311638712883, "loss": 0.003542384598404169 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 100, "global_step": 550, "compiler_max_steps": 24, "elapsed_sec": 1091.986, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 4.5789445721311495e-05, "op_loss": 0.23726259171962738, "arg_loss": 0.5636007189750671, "state_loss": 0.4535157382488251, "state_train_accuracy": 0.9618320465087891, "executor_loss": 3.182007074356079, "loss": 6.127371311187744 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 150, "global_step": 600, "compiler_max_steps": 24, "elapsed_sec": 1265.03, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 2.2217316654860042e-05, "op_loss": 7.518694360442169e-07, "arg_loss": 1.5298413927666843e-05, "state_loss": 0.00016169641457963735, "state_train_accuracy": 1.0, "executor_loss": 0.00030310172587633133, "loss": 0.0006156129529699683 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 200, "global_step": 650, "compiler_max_steps": 24, "elapsed_sec": 1438.714, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 1.5303367035812698e-05, "op_loss": 2.0656412743846886e-05, "arg_loss": 0.0001202040511998348, "state_loss": 0.0009623774676583707, "state_train_accuracy": 1.0, "executor_loss": 0.0024089086800813675, "loss": 0.003933972213417292 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 250, "global_step": 700, "compiler_max_steps": 24, "elapsed_sec": 1612.577, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 3.680487861856818e-05, "op_loss": 0.0008808095008134842, "arg_loss": 0.005063225049525499, "state_loss": 0.05041664466261864, "state_train_accuracy": 1.0, "executor_loss": 0.09213916212320328, "loss": 0.16383673250675201 }, { "run": "main_expand_copy_s750", "arm": "expand_copy", "stage": "stage3_max24", "stage_idx": 3, "local_step": 300, "global_step": 750, "compiler_max_steps": 24, "elapsed_sec": 1786.38, "expansion_event": "expand_16_to_24_copy_last", "init_loss": 4.45681216660887e-05, "op_loss": 0.28788819909095764, "arg_loss": 0.2507905662059784, "state_loss": 0.2743814289569855, "state_train_accuracy": 0.9674796462059021, "executor_loss": 1.0466102361679077, "loss": 2.612220525741577 } ], "metrics": [ { "stage": "stage1_max8", "split": "standard_L8", "n": 64, "max_steps": 8, "executor_accuracy": 1.0, "executor_target_mass": 0.9801869541406631, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 300, "stage_idx": 1 }, { "stage": "stage1_max8", "split": "paraphrase_L8", "n": 64, "max_steps": 8, "executor_accuracy": 0.984375, "executor_target_mass": 0.9692972674965858, "direct_accuracy": NaN, "init_accuracy": 0.984375, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 0.984375, "state_accuracy": 0.984375, "state_all_exact": 0.984375, "state_prefix_fraction": 0.984375, "state_mean_correct_prefix": 7.875, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 300, "stage_idx": 1 }, { "stage": "stage1_max8", "split": "heldout_L8", "n": 64, "max_steps": 8, "executor_accuracy": 0.75, "executor_target_mass": 0.7311536893248558, "direct_accuracy": NaN, "init_accuracy": 0.96875, "op_accuracy": 0.94140625, "arg_accuracy": 0.939453125, "program_exact": 0.734375, "state_accuracy": 0.845703125, "state_all_exact": 0.734375, "state_prefix_fraction": 0.841796875, "state_mean_correct_prefix": 6.734375, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 300, "stage_idx": 1 }, { "stage": "stage1_max8", "split": "paired_L8", "n": 64, "max_steps": 8, "executor_accuracy": 1.0, "executor_target_mass": 0.9977330192923546, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "executor_pair_answer_consistency": 1.0, "executor_pair_both_correct": 1.0, "compiler_pair_program_consistency": 1.0, "compiler_pair_state_consistency": 1.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 300, "stage_idx": 1 }, { "stage": "stage1_max8", "split": "paired_heldout_L8", "n": 64, "max_steps": 8, "executor_accuracy": 0.828125, "executor_target_mass": 0.8322859779000282, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.966796875, "arg_accuracy": 0.943359375, "program_exact": 0.828125, "state_accuracy": 0.91796875, "state_all_exact": 0.828125, "state_prefix_fraction": 0.91796875, "state_mean_correct_prefix": 7.34375, "executor_pair_answer_consistency": 0.65625, "executor_pair_both_correct": 0.65625, "compiler_pair_program_consistency": 0.65625, "compiler_pair_state_consistency": 0.65625, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 300, "stage_idx": 1 }, { "stage": "stage2_max16", "split": "standard_L8", "n": 64, "max_steps": 16, "executor_accuracy": 1.0, "executor_target_mass": 0.9989165365695953, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paraphrase_L8", "n": 64, "max_steps": 16, "executor_accuracy": 1.0, "executor_target_mass": 0.9994852840900421, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "heldout_L8", "n": 64, "max_steps": 16, "executor_accuracy": 0.90625, "executor_target_mass": 0.9037625417113304, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9609375, "arg_accuracy": 0.94921875, "program_exact": 0.90625, "state_accuracy": 0.939453125, "state_all_exact": 0.90625, "state_prefix_fraction": 0.9375, "state_mean_correct_prefix": 7.5, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paired_L8", "n": 64, "max_steps": 16, "executor_accuracy": 1.0, "executor_target_mass": 0.9994803443551064, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "executor_pair_answer_consistency": 1.0, "executor_pair_both_correct": 1.0, "compiler_pair_program_consistency": 1.0, "compiler_pair_state_consistency": 1.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paired_heldout_L8", "n": 64, "max_steps": 16, "executor_accuracy": 0.953125, "executor_target_mass": 0.9521941840648651, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.990234375, "arg_accuracy": 0.98046875, "program_exact": 0.953125, "state_accuracy": 0.98046875, "state_all_exact": 0.953125, "state_prefix_fraction": 0.978515625, "state_mean_correct_prefix": 7.828125, "executor_pair_answer_consistency": 0.90625, "executor_pair_both_correct": 0.90625, "compiler_pair_program_consistency": 0.90625, "compiler_pair_state_consistency": 0.90625, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "standard_L16", "n": 64, "max_steps": 16, "executor_accuracy": 0.953125, "executor_target_mass": 0.9425532966852188, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 0.9970703125, "program_exact": 0.953125, "state_accuracy": 0.9853515625, "state_all_exact": 0.953125, "state_prefix_fraction": 0.9853515625, "state_mean_correct_prefix": 15.765625, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paraphrase_L16", "n": 64, "max_steps": 16, "executor_accuracy": 0.9375, "executor_target_mass": 0.9120766595005989, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9970703125, "arg_accuracy": 0.9970703125, "program_exact": 0.9375, "state_accuracy": 0.9951171875, "state_all_exact": 0.9375, "state_prefix_fraction": 0.9951171875, "state_mean_correct_prefix": 15.921875, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "heldout_L16", "n": 64, "max_steps": 16, "executor_accuracy": 0.71875, "executor_target_mass": 0.6898384615778923, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9638671875, "arg_accuracy": 0.9482421875, "program_exact": 0.71875, "state_accuracy": 0.9228515625, "state_all_exact": 0.71875, "state_prefix_fraction": 0.921875, "state_mean_correct_prefix": 14.75, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paired_L16", "n": 64, "max_steps": 16, "executor_accuracy": 0.90625, "executor_target_mass": 0.8983785063028336, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 0.994140625, "program_exact": 0.90625, "state_accuracy": 0.9833984375, "state_all_exact": 0.90625, "state_prefix_fraction": 0.9833984375, "state_mean_correct_prefix": 15.734375, "executor_pair_answer_consistency": 0.8125, "executor_pair_both_correct": 0.8125, "compiler_pair_program_consistency": 0.8125, "compiler_pair_state_consistency": 0.8125, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage2_max16", "split": "paired_heldout_L16", "n": 64, "max_steps": 16, "executor_accuracy": 0.703125, "executor_target_mass": 0.7260855212807655, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.95703125, "arg_accuracy": 0.9423828125, "program_exact": 0.6875, "state_accuracy": 0.9287109375, "state_all_exact": 0.6875, "state_prefix_fraction": 0.927734375, "state_mean_correct_prefix": 14.84375, "executor_pair_answer_consistency": 0.46875, "executor_pair_both_correct": 0.46875, "compiler_pair_program_consistency": 0.4375, "compiler_pair_state_consistency": 0.4375, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 450, "stage_idx": 2 }, { "stage": "stage3_max24", "split": "standard_L8", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9993632212281227, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paraphrase_L8", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9984992295503616, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "heldout_L8", "n": 64, "max_steps": 24, "executor_accuracy": 0.796875, "executor_target_mass": 0.7922104969620705, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.95703125, "arg_accuracy": 0.990234375, "program_exact": 0.796875, "state_accuracy": 0.828125, "state_all_exact": 0.796875, "state_prefix_fraction": 0.828125, "state_mean_correct_prefix": 6.625, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_L8", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9997235015034676, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 8.0, "executor_pair_answer_consistency": 1.0, "executor_pair_both_correct": 1.0, "compiler_pair_program_consistency": 1.0, "compiler_pair_state_consistency": 1.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_heldout_L8", "n": 64, "max_steps": 24, "executor_accuracy": 0.875, "executor_target_mass": 0.876427911221981, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9765625, "arg_accuracy": 1.0, "program_exact": 0.875, "state_accuracy": 0.9140625, "state_all_exact": 0.875, "state_prefix_fraction": 0.9140625, "state_mean_correct_prefix": 7.3125, "executor_pair_answer_consistency": 0.75, "executor_pair_both_correct": 0.75, "compiler_pair_program_consistency": 0.75, "compiler_pair_state_consistency": 0.75, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "standard_L16", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9987226203083992, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 16.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paraphrase_L16", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9989713132381439, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 16.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "heldout_L16", "n": 64, "max_steps": 24, "executor_accuracy": 0.734375, "executor_target_mass": 0.7283697724342346, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9775390625, "arg_accuracy": 0.994140625, "program_exact": 0.734375, "state_accuracy": 0.779296875, "state_all_exact": 0.734375, "state_prefix_fraction": 0.779296875, "state_mean_correct_prefix": 12.46875, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_L16", "n": 64, "max_steps": 24, "executor_accuracy": 1.0, "executor_target_mass": 0.9991404265165329, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 1.0, "arg_accuracy": 1.0, "program_exact": 1.0, "state_accuracy": 1.0, "state_all_exact": 1.0, "state_prefix_fraction": 1.0, "state_mean_correct_prefix": 16.0, "executor_pair_answer_consistency": 1.0, "executor_pair_both_correct": 1.0, "compiler_pair_program_consistency": 1.0, "compiler_pair_state_consistency": 1.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_heldout_L16", "n": 64, "max_steps": 24, "executor_accuracy": 0.84375, "executor_target_mass": 0.8359821736812592, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.98046875, "arg_accuracy": 0.9814453125, "program_exact": 0.84375, "state_accuracy": 0.8896484375, "state_all_exact": 0.84375, "state_prefix_fraction": 0.8896484375, "state_mean_correct_prefix": 14.234375, "executor_pair_answer_consistency": 0.6875, "executor_pair_both_correct": 0.6875, "compiler_pair_program_consistency": 0.6875, "compiler_pair_state_consistency": 0.6875, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "standard_L24", "n": 64, "max_steps": 24, "executor_accuracy": 0.015625, "executor_target_mass": 0.010077128419652581, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.912109375, "arg_accuracy": 0.8645833333333334, "program_exact": 0.0, "state_accuracy": 0.859375, "state_all_exact": 0.0, "state_prefix_fraction": 0.8587239583333334, "state_mean_correct_prefix": 20.609375, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paraphrase_L24", "n": 64, "max_steps": 24, "executor_accuracy": 0.75, "executor_target_mass": 0.6844136640429497, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9850260416666666, "arg_accuracy": 0.974609375, "program_exact": 0.75, "state_accuracy": 0.9733072916666666, "state_all_exact": 0.75, "state_prefix_fraction": 0.9733072916666666, "state_mean_correct_prefix": 23.359375, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "heldout_L24", "n": 64, "max_steps": 24, "executor_accuracy": 0.515625, "executor_target_mass": 0.48618935979902744, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9635416666666666, "arg_accuracy": 0.974609375, "program_exact": 0.515625, "state_accuracy": 0.6861979166666666, "state_all_exact": 0.515625, "state_prefix_fraction": 0.6861979166666666, "state_mean_correct_prefix": 16.46875, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_L24", "n": 64, "max_steps": 24, "executor_accuracy": 0.296875, "executor_target_mass": 0.28734270483255386, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.9446614583333334, "arg_accuracy": 0.9140625, "program_exact": 0.296875, "state_accuracy": 0.9095052083333334, "state_all_exact": 0.296875, "state_prefix_fraction": 0.9095052083333334, "state_mean_correct_prefix": 21.828125, "executor_pair_answer_consistency": 0.09375, "executor_pair_both_correct": 0.0, "compiler_pair_program_consistency": 0.09375, "compiler_pair_state_consistency": 0.09375, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 }, { "stage": "stage3_max24", "split": "paired_heldout_L24", "n": 64, "max_steps": 24, "executor_accuracy": 0.25, "executor_target_mass": 0.2274046060629189, "direct_accuracy": NaN, "init_accuracy": 1.0, "op_accuracy": 0.939453125, "arg_accuracy": 0.9205729166666666, "program_exact": 0.234375, "state_accuracy": 0.7415364583333334, "state_all_exact": 0.234375, "state_prefix_fraction": 0.7408854166666666, "state_mean_correct_prefix": 17.78125, "executor_pair_answer_consistency": 0.0625, "executor_pair_both_correct": 0.03125, "compiler_pair_program_consistency": 0.0, "compiler_pair_state_consistency": 0.0, "run": "main_expand_copy_s750", "arm": "expand_copy", "global_step": 750, "stage_idx": 3 } ], "checkpoints": [ { "run": "main_expand_copy_s750", "tag": "stage1_max8_step00300", "stage": "stage1_max8", "global_step": 300, "compiler_max_steps": 8, "checkpoint_dir": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage1_max8_step00300", "model_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage1_max8_step00300/adapter", "heads_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage1_max8_step00300/heads.pt", "selection_split": "paired_L8", "selection_metric": "executor_pair_both_correct", "selection_value": 1.0, "executor_accuracy": 1.0, "program_exact": 1.0, "state_prefix_fraction": 1.0 }, { "run": "main_expand_copy_s750", "tag": "stage2_max16_step00450", "stage": "stage2_max16", "global_step": 450, "compiler_max_steps": 16, "checkpoint_dir": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage2_max16_step00450", "model_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage2_max16_step00450/adapter", "heads_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage2_max16_step00450/heads.pt", "selection_split": "paired_L8", "selection_metric": "executor_pair_both_correct", "selection_value": 1.0, "executor_accuracy": 1.0, "program_exact": 1.0, "state_prefix_fraction": 1.0 }, { "run": "main_expand_copy_s750", "tag": "stage3_max24_step00750", "stage": "stage3_max24", "global_step": 750, "compiler_max_steps": 24, "checkpoint_dir": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage3_max24_step00750", "model_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage3_max24_step00750/adapter", "heads_checkpoint": "/workspace/large_artifacts/qwen_structural_compiler_attribution_ablation/checkpoints/main_expand_copy_s750/stage3_max24_step00750/heads.pt", "selection_split": "paired_L24", "selection_metric": "executor_pair_both_correct", "selection_value": 0.0, "executor_accuracy": 0.296875, "program_exact": 0.296875, "state_prefix_fraction": 0.9095052083333334 } ], "elapsed_sec": 1914.242 }