{ "stage": "canonical_v5_corrected_dense_moe_page_multitask", "epoch": 3, "epoch_progress": 3.0, "epoch_completed": false, "global_step": 1914, "llm_name": "", "resume_from_checkpoint": "", "dense_initialization": { "mode": "dense_checkpoint_expand_moe_v1", "checkpoint_dir": "", "checkpoint_manifest": { "format_version": 2, "created_utc": "2026-08-31T06:26:34.597677+00:00", "base_model": "", "moe_signature": { "enabled": false, "routing_grouping": "script3", "language_groups": [ "latin", "greek", "han" ], "language_experts": 3, "shared_experts": 1, "layer_indices": [ 1, 5, 9, 13, 17, 21 ] }, "atomic_ocr_markers": [ "", "" ], "atomic_ocr_marker_ids": { "": 248077, "": 248078 }, "full_model_state_file": "full_model_state.pt", "training_state_included": true }, "checkpoint_metadata_path": "", "run_config_checked": true, "compared_contract": { "dataset": "page_multitask", "expected_sources": [ "BRESSAY", "Bentham", "HWDB2.0", "IAM", "ICFHR2016", "Images_Courriers", "ScribbleLens" ], "image_resize_policy": "area_ratio", "image_min_pixels": 589824, "image_max_pixels": 1572864, "image_max_aspect_ratio": 200.0, "max_text_length": 2560, "eval_max_new_tokens": 1024, "atomic_ocr_markers": [ "", "" ], "page_htr_eos_weight": 4.0, "token_ul_coef": 0.0, "token_ul_page_fraction": 0.0, "sampling_mode": "full_pass" }, "waived_contract_keys": [ { "key": "expected_split_records", "reason": "checkpoint_serialized_all_split_counts_as_null" } ], "source_scope_transfer": { "policy": "exact", "checkpoint_sources": [ "BRESSAY", "Bentham", "HWDB2.0", "IAM", "ICFHR2016", "Images_Courriers", "ScribbleLens" ], "active_sources": [ "BRESSAY", "Bentham", "HWDB2.0", "IAM", "ICFHR2016", "Images_Courriers", "ScribbleLens" ] }, "full_model_state_file": "", "full_model_state_bytes": 1705688804, "full_model_state_sha256": "7866b36b4d3cb9b3ce39a4dcd225ec51d2fde75a664ab1ad2260b4292b930388", "moe_transfer": { "checked": true, "layer_path": "model.language_model.layers", "layers": [ { "layer": 1, "max_abs_diff": 0.00048828125, "mean_abs_diff": 2.0134884834988043e-05, "mean_relative_diff": 0.0007749682408757508, "routed_experts": 4, "shared_experts": 1 }, { "layer": 5, "max_abs_diff": 0.0009765625, "mean_abs_diff": 3.1044164643390104e-05, "mean_relative_diff": 0.0011407412821426988, "routed_experts": 4, "shared_experts": 1 }, { "layer": 9, "max_abs_diff": 0.00048828125, "mean_abs_diff": 1.4513420865114313e-05, "mean_relative_diff": 0.0008208884974010289, "routed_experts": 4, "shared_experts": 1 }, { "layer": 13, "max_abs_diff": 0.00048828125, "mean_abs_diff": 1.9256414816481993e-05, "mean_relative_diff": 0.0008550191414542496, "routed_experts": 4, "shared_experts": 1 }, { "layer": 17, "max_abs_diff": 0.0009765625, "mean_abs_diff": 2.7810845494968817e-05, "mean_relative_diff": 0.0008891209727153182, "routed_experts": 4, "shared_experts": 1 }, { "layer": 21, "max_abs_diff": 0.0009765625, "mean_abs_diff": 2.683823367988225e-05, "mean_relative_diff": 0.0006347740418277681, "routed_experts": 4, "shared_experts": 1 } ], "tolerance": { "max_abs": 0.05, "mean_relative": 0.02 } } }, "train_metrics": { "train_loss": 0.5901720113664641, "train_lm_loss": 0.579147330208922, "train_moe_load_balance_raw_loss": 1.0158756909965416, "train_moe_load_balance_loss": 0.010158756687544316, "train_moe_router_language_loss": 0.0, "train_moe_lam_sparse_raw_loss": 0.07576503753092023, "train_moe_lam_sparse_loss": 0.0007576503584921332, "train_batches": 5097, "supervised_label_tokens": 1831248, "sample_exposure_epoch": { "draws": 20388, "unique_records": 20388, "duplicate_draws": 0, "unique_pages": 4393, "by_source": { "BRESSAY": 3726, "Bentham": 1842, "HWDB2.0": 5513, "IAM": 2988, "ICFHR2016": 1745, "Images_Courriers": 3690, "ScribbleLens": 884 }, "by_task": { "line_count": 4393, "line_range_htr": 7209, "page_htr": 4393, "phrase_line_lookup": 4393 }, "by_source_task": { "BRESSAY": { "line_count": 647, "line_range_htr": 1785, "page_htr": 647, "phrase_line_lookup": 647 }, "Bentham": { "line_count": 350, "line_range_htr": 792, "page_htr": 350, "phrase_line_lookup": 350 }, "HWDB2.0": { "line_count": 1377, "line_range_htr": 1382, "page_htr": 1377, "phrase_line_lookup": 1377 }, "IAM": { "line_count": 747, "line_range_htr": 747, "page_htr": 747, "phrase_line_lookup": 747 }, "ICFHR2016": { "line_count": 350, "line_range_htr": 695, "page_htr": 350, "phrase_line_lookup": 350 }, "Images_Courriers": { "line_count": 758, "line_range_htr": 1416, "page_htr": 758, "phrase_line_lookup": 758 }, "ScribbleLens": { "line_count": 164, "line_range_htr": 392, "page_htr": 164, "phrase_line_lookup": 164 } }, "unique_records_by_source": { "BRESSAY": 3726, "Bentham": 1842, "HWDB2.0": 5513, "IAM": 2988, "ICFHR2016": 1745, "Images_Courriers": 3690, "ScribbleLens": 884 }, "unique_pages_by_source": { "BRESSAY": 647, "Bentham": 350, "HWDB2.0": 1377, "IAM": 747, "ICFHR2016": 350, "Images_Courriers": 758, "ScribbleLens": 164 } }, "sample_exposure_cumulative": { "draws": 61164, "unique_records": 20388, "duplicate_draws": 40776, "unique_pages": 4393, "by_source": { "BRESSAY": 11178, "Bentham": 5526, "HWDB2.0": 16539, "IAM": 8964, "ICFHR2016": 5235, "Images_Courriers": 11070, "ScribbleLens": 2652 }, "by_task": { "line_count": 13179, "line_range_htr": 21627, "page_htr": 13179, "phrase_line_lookup": 13179 }, "by_source_task": { "BRESSAY": { "line_count": 1941, "line_range_htr": 5355, "page_htr": 1941, "phrase_line_lookup": 1941 }, "Bentham": { "line_count": 1050, "line_range_htr": 2376, "page_htr": 1050, "phrase_line_lookup": 1050 }, "HWDB2.0": { "line_count": 4131, "line_range_htr": 4146, "page_htr": 4131, "phrase_line_lookup": 4131 }, "IAM": { "line_count": 2241, "line_range_htr": 2241, "page_htr": 2241, "phrase_line_lookup": 2241 }, "ICFHR2016": { "line_count": 1050, "line_range_htr": 2085, "page_htr": 1050, "phrase_line_lookup": 1050 }, "Images_Courriers": { "line_count": 2274, "line_range_htr": 4248, "page_htr": 2274, "phrase_line_lookup": 2274 }, "ScribbleLens": { "line_count": 492, "line_range_htr": 1176, "page_htr": 492, "phrase_line_lookup": 492 } }, "unique_records_by_source": { "BRESSAY": 3726, "Bentham": 1842, "HWDB2.0": 5513, "IAM": 2988, "ICFHR2016": 1745, "Images_Courriers": 3690, "ScribbleLens": 884 }, "unique_pages_by_source": { "BRESSAY": 647, "Bentham": 350, "HWDB2.0": 1377, "IAM": 747, "ICFHR2016": 350, "Images_Courriers": 758, "ScribbleLens": 164 } } }, "val_metrics": { "cer_generate": 0.16438965258561158, "wer_generate": 0.33299821549752434, "num_samples": 962, "safety_cer_generate_diagnostic": 0.15750271999682047, "loop_guard_trigger_rate": 0.028066528066528068, "by_source": { "BRESSAY": { "cer_generate": 0.20291806036232865, "wer_generate": 0.3477519644157897, "num_samples": 154 }, "Bentham": { "cer_generate": 0.1773907135467334, "wer_generate": 0.4670343876583642, "num_samples": 50 }, "HWDB2.0": { "cer_generate": 0.031286644544796786, "wer_generate": 0.3711730306157551, "num_samples": 300 }, "IAM": { "cer_generate": 0.020206642421906985, "wer_generate": 0.06037131027113664, "num_samples": 116 }, "ICFHR2016": { "cer_generate": 0.4769110835970101, "wer_generate": 0.8564913347522043, "num_samples": 50 }, "Images_Courriers": { "cer_generate": 0.08557112516173132, "wer_generate": 0.1783308596871789, "num_samples": 278 }, "ScribbleLens": { "cer_generate": 0.4741920423302381, "wer_generate": 0.87206106870229, "num_samples": 14 } }, "safety_cer_by_source_diagnostic": { "BRESSAY": 0.20162264844067207, "Bentham": 0.1496116489862726, "HWDB2.0": 0.031286644544796786, "IAM": 0.020206642421906985, "ICFHR2016": 0.4769110835970101, "Images_Courriers": 0.07008153699680199, "ScribbleLens": 0.4393318462416351 }, "macro_source_cer_diagnostic": 0.20978233028067791, "macro_page_cer_diagnostic": 0.10694783478550064, "by_task": { "page_htr": { "cer_generate": 0.16438965258561158, "wer_generate": 0.33299821549752434, "num_samples": 962 } }, "by_language_group": { "all": { "cer_generate": 0.16438965258561158, "wer_generate": 0.33299821549752434, "num_samples": 962 } }, "generation_diagnostics": { "num_samples": 962, "output_tokens": { "count": 962, "mean": 277.5686070686071, "p50": 190.0, "p95": 754.0, "p99": 1024.0, "max": 1024.0 }, "reference_tokens": { "count": 962, "mean": 260.11538461538464, "p50": 186.0, "p95": 672.9499999999999, "p99": 806.39, "max": 944.0 }, "output_reference_token_ratio": { "count": 962, "mean": 1.0538785060645548, "p50": 1.0136054421768708, "p95": 1.1400505503262113, "p99": 2.213666242038215, "max": 5.333333333333333 }, "prediction_chars": { "count": 962, "mean": 860.0436590436591, "p50": 501.5, "p95": 2838.95, "p99": 3455.29, "max": 4349.0 }, "reference_chars": { "count": 962, "mean": 855.5197505197506, "p50": 507.0, "p95": 2833.049999999999, "p99": 3482.68, "max": 4298.0 }, "max_new_tokens_hit_rate": 0.02702702702702703, "eos_emitted_rate": 0.972972972972973, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.028066528066528068, "repetition_3gram_fraction": { "count": 962, "mean": 0.042803856253832664, "p50": 0.017991044937152756, "p95": 0.10372954615622096, "p99": 0.7909393346379647, "max": 0.9363992172211351 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.8056133056133056, "marker_reference_counts": { "": 576, "": 297 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 962, "mean": 1.8000839774241226, "p50": 1.1455812877502467, "p95": 5.233787497999401, "p99": 5.325606547574771, "max": 6.0885678070008 }, "generation_seconds_per_sample": { "count": 962, "mean": 1.8000839774241226, "p50": 1.1455812877502467, "p95": 5.233787497999401, "p99": 5.325606547574771, "max": 6.0885678070008 }, "generation_wall_seconds_total": 1731.680786282006, "generation_pages_per_second": 0.5555296378066629, "generated_tokens_per_second": 154.19758775132323 }, "evaluation_wall_seconds": 1747.2859408959994, "diagnostics_by_source": { "BRESSAY": { "num_samples": 154, "output_tokens": { "count": 154, "mean": 685.2857142857143, "p50": 655.5, "p95": 1024.0, "p99": 1024.0, "max": 1024.0 }, "reference_tokens": { "count": 154, "mean": 646.6883116883117, "p50": 636.5, "p95": 814.4, "p99": 925.35, "max": 944.0 }, "output_reference_token_ratio": { "count": 154, "mean": 1.0582472392281992, "p50": 1.0112063214077225, "p95": 1.4542156304832359, "p99": 1.654286506305147, "max": 1.7038269550748752 }, "prediction_chars": { "count": 154, "mean": 2685.525974025974, "p50": 2648.0, "p95": 3469.85, "p99": 4216.37, "max": 4349.0 }, "reference_chars": { "count": 154, "mean": 2708.8311688311687, "p50": 2663.5, "p95": 3513.4, "p99": 3693.27, "max": 4298.0 }, "max_new_tokens_hit_rate": 0.09090909090909091, "eos_emitted_rate": 0.9090909090909091, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.1038961038961039, "repetition_3gram_fraction": { "count": 154, "mean": 0.08230364542639139, "p50": 0.027308844865693127, "p95": 0.44138943248532264, "p99": 0.793170254403131, "max": 0.821917808219178 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.7012987012987013, "marker_reference_counts": { "": 560, "": 57 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 154, "mean": 4.209619012707757, "p50": 3.977695644999585, "p95": 5.30480367291184, "p99": 5.341092155499609, "max": 5.341092155499609 }, "generation_seconds_per_sample": { "count": 154, "mean": 4.209619012707757, "p50": 3.977695644999585, "p95": 5.30480367291184, "p99": 5.341092155499609, "max": 5.341092155499609 }, "generation_wall_seconds_total": 648.2813279569946, "generation_pages_per_second": 0.23755118859480093, "generated_tokens_per_second": 162.79043595560856 }, "Bentham": { "num_samples": 50, "output_tokens": { "count": 50, "mean": 422.12, "p50": 332.5, "p95": 1024.0, "p99": 1024.0, "max": 1024.0 }, "reference_tokens": { "count": 50, "mean": 330.94, "p50": 320.5, "p95": 481.0, "p99": 536.14, "max": 543.0 }, "output_reference_token_ratio": { "count": 50, "mean": 1.2766760861172306, "p50": 1.0, "p95": 3.2397365295324465, "p99": 3.6336356432156376, "max": 3.696750902527076 }, "prediction_chars": { "count": 50, "mean": 1358.92, "p50": 1321.5, "p95": 2073.499999999999, "p99": 2334.75, "max": 2347.0 }, "reference_chars": { "count": 50, "mean": 1310.18, "p50": 1252.5, "p95": 1860.35, "p99": 2225.5399999999995, "max": 2350.0 }, "max_new_tokens_hit_rate": 0.14, "eos_emitted_rate": 0.86, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.14, "repetition_3gram_fraction": { "count": 50, "mean": 0.15351462873620134, "p50": 0.0354426409472281, "p95": 0.8212818003913894, "p99": 0.8724266144814089, "max": 0.9021526418786693 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.24, "marker_reference_counts": { "": 0, "": 4 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 50, "mean": 3.4135301617999723, "p50": 2.6635789757492603, "p95": 5.2451707629998054, "p99": 5.2451707629998054, "max": 5.2451707629998054 }, "generation_seconds_per_sample": { "count": 50, "mean": 3.4135301617999723, "p50": 2.6635789757492603, "p95": 5.2451707629998054, "p99": 5.2451707629998054, "max": 5.2451707629998054 }, "generation_wall_seconds_total": 170.6765080899986, "generation_pages_per_second": 0.29295185705132154, "generated_tokens_per_second": 123.66083789850386 }, "HWDB2.0": { "num_samples": 300, "output_tokens": { "count": 300, "mean": 167.40666666666667, "p50": 167.0, "p95": 201.0, "p99": 213.0, "max": 214.0 }, "reference_tokens": { "count": 300, "mean": 165.7, "p50": 164.5, "p95": 199.0, "p99": 209.01999999999998, "max": 213.0 }, "output_reference_token_ratio": { "count": 300, "mean": 1.0106741074467964, "p50": 1.0116959064327486, "p95": 1.038722245496439, "p99": 1.0508463221219577, "max": 1.0793650793650793 }, "prediction_chars": { "count": 300, "mean": 272.2966666666667, "p50": 269.0, "p95": 313.0, "p99": 326.0, "max": 332.0 }, "reference_chars": { "count": 300, "mean": 276.35333333333335, "p50": 275.0, "p95": 317.15000000000003, "p99": 328.03999999999996, "max": 343.0 }, "max_new_tokens_hit_rate": 0.0, "eos_emitted_rate": 1.0, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.0, "repetition_3gram_fraction": { "count": 300, "mean": 0.01982862275667522, "p50": 0.01365203615692856, "p95": 0.051183712121212144, "p99": 0.05883968972204266, "max": 0.07017543859649122 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.9633333333333334, "marker_reference_counts": { "": 0, "": 236 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 300, "mean": 1.0133197663800093, "p50": 0.9922402704996784, "p95": 1.1455812877502467, "p99": 1.1506245174996366, "max": 1.1506245174996366 }, "generation_seconds_per_sample": { "count": 300, "mean": 1.0133197663800093, "p50": 0.9922402704996784, "p95": 1.1455812877502467, "p99": 1.1506245174996366, "max": 1.1506245174996366 }, "generation_wall_seconds_total": 303.9959299140028, "generation_pages_per_second": 0.986855317717138, "generated_tokens_per_second": 165.20615922130034 }, "IAM": { "num_samples": 116, "output_tokens": { "count": 116, "mean": 98.0, "p50": 97.0, "p95": 127.0, "p99": 155.04999999999993, "max": 157.0 }, "reference_tokens": { "count": 116, "mean": 95.91379310344827, "p50": 94.5, "p95": 123.75, "p99": 151.04999999999993, "max": 155.0 }, "output_reference_token_ratio": { "count": 116, "mean": 1.0228742961042576, "p50": 1.0209429824561402, "p95": 1.0460884353741497, "p99": 1.093605025773196, "max": 1.125 }, "prediction_chars": { "count": 116, "mean": 364.2844827586207, "p50": 358.5, "p95": 501.25, "p99": 582.5499999999997, "max": 610.0 }, "reference_chars": { "count": 116, "mean": 363.9396551724138, "p50": 358.5, "p95": 497.5, "p99": 580.8499999999998, "max": 610.0 }, "max_new_tokens_hit_rate": 0.0, "eos_emitted_rate": 1.0, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.0, "repetition_3gram_fraction": { "count": 116, "mean": 0.005795495058881031, "p50": 0.0, "p95": 0.029751131221719462, "p99": 0.050506072874493886, "max": 0.11538461538461542 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 1.0, "marker_reference_counts": { "": 0, "": 0 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 116, "mean": 0.6505729309310482, "p50": 0.6363210999998046, "p95": 0.8624965879998854, "p99": 0.8652169607494216, "max": 0.8652169607494216 }, "generation_seconds_per_sample": { "count": 116, "mean": 0.6505729309310482, "p50": 0.6363210999998046, "p95": 0.8624965879998854, "p99": 0.8652169607494216, "max": 0.8652169607494216 }, "generation_wall_seconds_total": 75.46645998800159, "generation_pages_per_second": 1.5371066831337106, "generated_tokens_per_second": 150.63645494710363 }, "ICFHR2016": { "num_samples": 50, "output_tokens": { "count": 50, "mean": 228.18, "p50": 230.5, "p95": 270.0, "p99": 661.3999999999985, "max": 1024.0 }, "reference_tokens": { "count": 50, "mean": 192.18, "p50": 209.5, "p95": 254.1, "p99": 260.59, "max": 265.0 }, "output_reference_token_ratio": { "count": 50, "mean": 1.214897954381385, "p50": 1.0978468899521532, "p95": 1.4179384203480587, "p99": 3.5281818181818108, "max": 5.333333333333333 }, "prediction_chars": { "count": 50, "mean": 496.52, "p50": 494.5, "p95": 596.1, "p99": 1514.1199999999967, "max": 2353.0 }, "reference_chars": { "count": 50, "mean": 459.26, "p50": 500.5, "p95": 603.5, "p99": 624.02, "max": 625.0 }, "max_new_tokens_hit_rate": 0.02, "eos_emitted_rate": 0.98, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.0, "repetition_3gram_fraction": { "count": 50, "mean": 0.06765920660542266, "p50": 0.040066703526012604, "p95": 0.15343333333333317, "p99": 0.5539309761517781, "max": 0.8825831702544031 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.6, "marker_reference_counts": { "": 16, "": 0 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 50, "mean": 1.6397333451399028, "p50": 1.3918661629995768, "p95": 5.401545224499387, "p99": 5.401545224499387, "max": 5.401545224499387 }, "generation_seconds_per_sample": { "count": 50, "mean": 1.6397333451399028, "p50": 1.3918661629995768, "p95": 5.401545224499387, "p99": 5.401545224499387, "max": 5.401545224499387 }, "generation_wall_seconds_total": 81.98666725699513, "generation_pages_per_second": 0.6098552566269119, "generated_tokens_per_second": 139.15677245712874 }, "Images_Courriers": { "num_samples": 278, "output_tokens": { "count": 278, "mean": 213.79856115107913, "p50": 205.0, "p95": 289.44999999999993, "p99": 383.46000000000004, "max": 1024.0 }, "reference_tokens": { "count": 278, "mean": 206.23381294964028, "p50": 201.0, "p95": 289.44999999999993, "p99": 337.7300000000009, "max": 450.0 }, "output_reference_token_ratio": { "count": 278, "mean": 1.0309230199193677, "p50": 1.013667496886675, "p95": 1.0607790801851575, "p99": 1.239687500000001, "max": 4.19672131147541 }, "prediction_chars": { "count": 278, "mean": 633.8597122302158, "p50": 605.5, "p95": 891.3, "p99": 1262.4400000000023, "max": 3107.0 }, "reference_chars": { "count": 278, "mean": 621.9208633093525, "p50": 604.5, "p95": 873.0999999999997, "p99": 1076.3100000000036, "max": 1486.0 }, "max_new_tokens_hit_rate": 0.007194244604316547, "eos_emitted_rate": 0.9928057553956835, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.007194244604316547, "repetition_3gram_fraction": { "count": 278, "mean": 0.02890799216476402, "p50": 0.017416582406471204, "p95": 0.06890628879509099, "p99": 0.16417016953902247, "max": 0.7681017612524462 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.7769784172661871, "marker_reference_counts": { "": 0, "": 0 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 278, "mean": 1.4111926457698238, "p50": 1.2364898257501409, "p95": 1.9682605375001003, "p99": 5.259757931499735, "max": 5.259757931499735 }, "generation_seconds_per_sample": { "count": 278, "mean": 1.4111926457698238, "p50": 1.2364898257501409, "p95": 1.9682605375001003, "p99": 5.259757931499735, "max": 5.259757931499735 }, "generation_wall_seconds_total": 392.31155552401106, "generation_pages_per_second": 0.7086204729011233, "generated_tokens_per_second": 151.50203750845742 }, "ScribbleLens": { "num_samples": 14, "output_tokens": { "count": 14, "mean": 567.5714285714286, "p50": 569.0, "p95": 1024.0, "p99": 1024.0, "max": 1024.0 }, "reference_tokens": { "count": 14, "mean": 451.14285714285717, "p50": 480.5, "p95": 590.7, "p99": 591.74, "max": 592.0 }, "output_reference_token_ratio": { "count": 14, "mean": 1.2735782748793814, "p50": 1.1085131559332142, "p95": 2.3024935067671866, "p99": 2.4932530437851987, "max": 2.5409429280397022 }, "prediction_chars": { "count": 14, "mean": 1490.0, "p50": 1583.5, "p95": 2222.3999999999996, "p99": 2442.8799999999997, "max": 2498.0 }, "reference_chars": { "count": 14, "mean": 1382.9285714285713, "p50": 1489.0, "p95": 1795.55, "p99": 1802.31, "max": 1804.0 }, "max_new_tokens_hit_rate": 0.14285714285714285, "eos_emitted_rate": 0.8571428571428571, "empty_output_rate": 0.0, "loop_guard_trigger_rate": 0.14285714285714285, "repetition_3gram_fraction": { "count": 14, "mean": 0.19904074841459948, "p50": 0.056307386702582396, "p95": 0.9287671232876713, "p99": 0.9348727984344423, "max": 0.9363992172211351 }, "region_count_exact_rate": 1.0, "line_count_exact_rate": 0.2857142857142857, "marker_reference_counts": { "": 0, "": 0 }, "marker_prediction_counts": { "": 0, "": 0 }, "generation_seconds_per_page": { "count": 14, "mean": 4.2115955394287345, "p50": 3.213291670250328, "p95": 6.0885678070008, "p99": 6.0885678070008, "max": 6.0885678070008 }, "generation_seconds_per_sample": { "count": 14, "mean": 4.2115955394287345, "p50": 3.213291670250328, "p95": 6.0885678070008, "p99": 6.0885678070008, "max": 6.0885678070008 }, "generation_wall_seconds_total": 58.96233755200228, "generation_pages_per_second": 0.2374397044155957, "generated_tokens_per_second": 134.76399223473737 } } }, "checkpoint_dir": "", "architecture": "qwen3_5_vlm_script_grouped_supervised_mlp_moe", "dataset_config": { "dataset_name": "page_multitask", "data_root": "", "train_json_path": "", "valid_json_path": "", "test_json_path": "", "image_root": "", "prompt_text": "", "expected_sources": [ "BRESSAY", "Bentham", "HWDB2.0", "IAM", "ICFHR2016", "Images_Courriers", "ScribbleLens" ], "excluded_sources": [ "Eparchos" ], "excluded_train_tasks": [], "excluded_valid_tasks": [], "structure_task_policy": "page_total_physical_line_count_v3", "image_resize_policy": "area_ratio", "image_resize_factor": 32, "image_max_aspect_ratio": 200.0, "image_min_pixels": 589824, "image_max_pixels": 1572864, "resolution_profile": "area_ratio_min589824_max1572864_ctx2560", "max_text_length": 2560, "processor_image_size": { "shortest_edge": 589824, "longest_edge": 1572864 }, "sampling_mode": "full_pass", "microbatch_invariant_token_loss": true, "training_terminal_id": 248044, "generation_stop_ids": [ 248044 ], "atomic_ocr_marker_ids": { "": 248077, "": 248078 }, "page_htr_eos_weight": 4.0, "token_ul_coef": 0.0, "token_ul_page_fraction": 0.0, "loop_guard": { "ngram": 3, "window": 96, "threshold": 0.6, "min_tokens": 128, "consecutive": 2 } }, "best_metric": 0.16438965258561158, "evaluations_no_improve": 0, "validation_has_started": true, "top_checkpoints": [ { "cer": 0.16947443202988768, "global_step": 638, "epoch_progress": 1.0, "path": "" } ], "learning_rate": 0.0, "loss_weighting": "optimizer_window_eos_weighted_ce_plus_token_ul_plus_router_balance_and_lam_sparse", "moe_signature": { "enabled": true, "ablation_mode": "sparsegen_shared_lamsparse_balance", "routing_grouping": "unsupervised4", "language_groups": [ "all" ], "language_experts": 4, "shared_experts": 1, "shared_mix": 0.25, "layer_indices": [ 1, 5, 9, 13, 17, 21 ], "routing_fn": "sparsegen", "router_ce_coef": 0.0, "router_ce_class_weighting": "none", "load_balance_coef": 0.01, "load_balance_warmup_fraction": 0.1, "load_balance_warmup_updates": 64, "lam_sparse_coef": 0.01, "lam_sparse_warmup_fraction": 0.1, "router_temperature": 1.0, "sparsegen_hidden": 128, "sparsegen_eps": 0.01, "expert_init": "copy", "expert_train_dtype": "match_base", "freeze_base": false }, "moe_mlp": { "routing_grouping": "unsupervised4", "routing_group_description": "four source-agnostic routed experts; no router supervision", "routing_fn": "sparsegen", "router_temperature": 1.0, "load_balance_coef": 0.01, "load_balance_warmup_fraction": 0.1, "load_balance_warmup_updates": 64, "router_ce_coef": 0.0, "lam_sparse_coef": 0.01, "lam_sparse_warmup_fraction": 0.1, "lam_sparse_warmup_updates": 64, "shared_mix": 0.25, "hard_top_k": false, "active_expert_budget": 0, "routed_top_k": 4, "freeze_base": false, "language_group_counts": {}, "language_group_class_weights": {}, "metadata_used_for_training_router_labels": false, "metadata_used_for_inference_routing": false } }