{ "format": "vimeml_public_evaluation_v4", "model_version": "4.2-epoch20-step1081942", "checkpoint_sha256": "0e83ea7417b9816b37183044468343b151f4e610da0764c31f1f780c5fa5796d", "coreml_manifest_sha256": { "stateless_int8": "0c3b6f3b809aaf3dd2f9ae18626c363f72dd34c59953dea48fe865ebbde9d6da", "kvchunk_int8": "fd4bbb6c31fa2f71b2c9be6ab5feb0bbb03b7f51216907c7ff271171d26d1b75" }, "training": { "epochs": 20, "updates": 1081942, "seed": 1219827974, "supervised_targets": 12104389994, "full_validation_lm_bpc_fp32": 3.0451515462768692, "full_validation_chat_bpc_fp32": 3.122059834972166, "selection": "best complete LM validation BPC (epoch 20), fixed before any IME scoring", "test_or_blind_used": false }, "fixed_pool_fp32_cuda": { "standard-development": { "azookey": { "cases": 728, "top1_correct": 453, "top1_accuracy": 0.6222527472527473, "top5_correct": 608, "top5_accuracy": 0.8351648351648352, "candidate_pool_covered": 630, "candidate_pool_recall": 0.8653846153846154, "covered_top1_accuracy": 0.719047619047619, "mean_min_cer": 0.05160664332397995, "mrr": 0.7097974489912511, "corrected_vs_azookey": 0, "regressed_vs_azookey": 0, "fallback_cases": 0 }, "lm_context_sum": { "cases": 728, "top1_correct": 523, "top1_accuracy": 0.7184065934065934, "top5_correct": 624, "top5_accuracy": 0.8571428571428571, "candidate_pool_covered": 630, "candidate_pool_recall": 0.8653846153846154, "covered_top1_accuracy": 0.8301587301587302, "mean_min_cer": 0.034705386175760286, "mrr": 0.7749688955732909, "corrected_vs_azookey": 110, "regressed_vs_azookey": 40, "fallback_cases": 0 }, "lm_no_context_sum": { "cases": 728, "top1_correct": 510, "top1_accuracy": 0.7005494505494505, "top5_correct": 621, "top5_accuracy": 0.853021978021978, "candidate_pool_covered": 630, "candidate_pool_recall": 0.8653846153846154, "covered_top1_accuracy": 0.8095238095238095, "mean_min_cer": 0.03638251311053943, "mrr": 0.7641914137449851, "corrected_vs_azookey": 108, "regressed_vs_azookey": 51, "fallback_cases": 0 } }, "legacy-regression": { "azookey": { "cases": 500, "top1_correct": 317, "top1_accuracy": 0.634, "top5_correct": 437, "top5_accuracy": 0.874, "candidate_pool_covered": 443, "candidate_pool_recall": 0.886, "covered_top1_accuracy": 0.7155756207674944, "mean_min_cer": 0.04623841144643269, "mrr": 0.7311617826617824, "corrected_vs_azookey": 0, "regressed_vs_azookey": 0, "fallback_cases": 0 }, "lm_context_sum": { "cases": 500, "top1_correct": 370, "top1_accuracy": 0.74, "top5_correct": 442, "top5_accuracy": 0.884, "candidate_pool_covered": 443, "candidate_pool_recall": 0.886, "covered_top1_accuracy": 0.835214446952596, "mean_min_cer": 0.028879998926345758, "mrr": 0.8080666666666665, "corrected_vs_azookey": 76, "regressed_vs_azookey": 23, "fallback_cases": 0 }, "lm_no_context_sum": { "cases": 500, "top1_correct": 369, "top1_accuracy": 0.738, "top5_correct": 442, "top5_accuracy": 0.884, "candidate_pool_covered": 443, "candidate_pool_recall": 0.886, "covered_top1_accuracy": 0.8329571106094809, "mean_min_cer": 0.028316756130444452, "mrr": 0.8037999999999997, "corrected_vs_azookey": 77, "regressed_vs_azookey": 25, "fallback_cases": 0 } }, "ajimee": { "azookey": { "cases": 200, "top1_correct": 87, "top1_accuracy": 0.435, "top5_correct": 143, "top5_accuracy": 0.715, "candidate_pool_covered": 162, "candidate_pool_recall": 0.81, "covered_top1_accuracy": 0.5370370370370371, "mean_min_cer": 0.0821630611981915, "mrr": 0.554232017982018, "corrected_vs_azookey": 0, "regressed_vs_azookey": 0, "fallback_cases": 0 }, "lm_context_sum": { "cases": 200, "top1_correct": 139, "top1_accuracy": 0.695, "top5_correct": 161, "top5_accuracy": 0.805, "candidate_pool_covered": 162, "candidate_pool_recall": 0.81, "covered_top1_accuracy": 0.8580246913580247, "mean_min_cer": 0.03520357489337187, "mrr": 0.7448809523809524, "corrected_vs_azookey": 58, "regressed_vs_azookey": 6, "fallback_cases": 0 }, "lm_no_context_sum": { "cases": 200, "top1_correct": 130, "top1_accuracy": 0.65, "top5_correct": 161, "top5_accuracy": 0.805, "candidate_pool_covered": 162, "candidate_pool_recall": 0.81, "covered_top1_accuracy": 0.8024691358024691, "mean_min_cer": 0.051505004672742855, "mrr": 0.7155833333333333, "corrected_vs_azookey": 52, "regressed_vs_azookey": 9, "fallback_cases": 0 } }, "historical-development": { "azookey": { "cases": 137, "top1_correct": 111, "top1_accuracy": 0.8102189781021898, "top5_correct": 130, "top5_accuracy": 0.948905109489051, "candidate_pool_covered": 136, "candidate_pool_recall": 0.9927007299270073, "covered_top1_accuracy": 0.8161764705882353, "mean_min_cer": 0.04746998743349109, "mrr": 0.8745925911071898, "corrected_vs_azookey": 0, "regressed_vs_azookey": 0, "fallback_cases": 0 }, "lm_context_sum": { "cases": 137, "top1_correct": 123, "top1_accuracy": 0.8978102189781022, "top5_correct": 136, "top5_accuracy": 0.9927007299270073, "candidate_pool_covered": 136, "candidate_pool_recall": 0.9927007299270073, "covered_top1_accuracy": 0.9044117647058824, "mean_min_cer": 0.02333226384321275, "mrr": 0.9388077858880779, "corrected_vs_azookey": 15, "regressed_vs_azookey": 3, "fallback_cases": 0 }, "lm_no_context_sum": { "cases": 137, "top1_correct": 124, "top1_accuracy": 0.9051094890510949, "top5_correct": 136, "top5_accuracy": 0.9927007299270073, "candidate_pool_covered": 136, "candidate_pool_recall": 0.9927007299270073, "covered_top1_accuracy": 0.9117647058823529, "mean_min_cer": 0.019706327735524817, "mrr": 0.9454987834549878, "corrected_vs_azookey": 16, "regressed_vs_azookey": 3, "fallback_cases": 0 } } }, "fixed_reference_725": { "plan_id": "81512f147d3749ef9a36e61470fce173", "role": "reference_test_after_consumption", "source_metrics": { "by_source": { "jmultiwoz": { "cases": 250, "top1_correct": 226, "top1_accuracy": 0.904, "top5_correct": 245, "top5_accuracy": 0.98, "candidate_pool_covered": 245, "candidate_pool_recall": 0.98, "covered_top1_accuracy": 0.9224489795918367, "mean_min_cer": 0.009039811881542249, "mrr": 0.9376666666666666, "corrected_vs_azookey": 35, "regressed_vs_azookey": 11, "fallback_cases": 0 }, "wrime": { "cases": 475, "top1_correct": 315, "top1_accuracy": 0.6631578947368421, "top5_correct": 398, "top5_accuracy": 0.8378947368421052, "candidate_pool_covered": 405, "candidate_pool_recall": 0.8526315789473684, "covered_top1_accuracy": 0.7777777777777778, "mean_min_cer": 0.05086094251484006, "mrr": 0.7354361661730081, "corrected_vs_azookey": 82, "regressed_vs_azookey": 32, "fallback_cases": 0 } }, "independent_micro": { "cases": 725, "top1_correct": 541, "top1_accuracy": 0.7462068965517241, "top5_correct": 643, "top5_accuracy": 0.886896551724138, "candidate_pool_covered": 650, "candidate_pool_recall": 0.896551724137931, "covered_top1_accuracy": 0.8323076923076923, "mean_min_cer": 0.03643986298611669, "mrr": 0.805170821515649, "corrected_vs_azookey": 117, "regressed_vs_azookey": 43, "fallback_cases": 0 }, "independent_equal_source_macro_top1": 0.783578947368421 }, "paired_vs_v21_fp32": { "v1_wrong_v2_correct": 40, "v1_correct_v2_wrong": 34, "both_correct": 501, "both_wrong": 150, "net_correct": 6, "exact_mcnemar_p_two_sided": 0.5613807442989819 }, "by_source_vs_v21": { "wrime": { "v1_wrong_v2_correct": 32, "v1_correct_v2_wrong": 26, "both_correct": 283, "both_wrong": 134, "net_correct": 6, "exact_mcnemar_p_two_sided": 0.5118423084381689 }, "jmultiwoz": { "v1_wrong_v2_correct": 8, "v1_correct_v2_wrong": 8, "both_correct": 218, "both_wrong": 16, "net_correct": 0, "exact_mcnemar_p_two_sided": 1 } } }, "coreml": { "top1": { "cuda": { "standard-development": 523, "legacy-regression": 370, "ajimee": 139, "historical-development": 123 }, "mac_pytorch": { "standard-development": 523, "legacy-regression": 370, "ajimee": 139, "historical-development": 123 }, "coreml_fp32": { "standard-development": 523, "legacy-regression": 370, "ajimee": 139, "historical-development": 123 }, "coreml_int8": { "standard-development": 522, "legacy-regression": 371, "ajimee": 139, "historical-development": 123 }, "kv_fp32": { "standard-development": 523, "legacy-regression": 370, "ajimee": 139, "historical-development": 123 }, "kv_int8": { "standard-development": 522, "legacy-regression": 371, "ajimee": 139, "historical-development": 123 }, "kvchunk_fp32": { "standard-development": 523, "legacy-regression": 370, "ajimee": 139, "historical-development": 123 }, "kvchunk_int8": { "standard-development": 522, "legacy-regression": 371, "ajimee": 139, "historical-development": 123 }, "v21_coreml_fp32": { "standard-development": 505, "legacy-regression": 374, "ajimee": 144, "historical-development": 122 }, "v21_coreml_int8": { "standard-development": 506, "legacy-regression": 375, "ajimee": 144, "historical-development": 122 } }, "checks": { "1_mac_pytorch_fixtures_within_3e-4": true, "1_mac_pytorch_top1_equals_recorded_cuda": true, "1_mac_pytorch_orders_equal_cuda_all_methods": true, "2_fp32_stateless_and_kv_converted": true, "3_fp32_stateless_alignment_passed": true, "3_fp32_kv_alignment_passed": true, "3_coreml_fp32_top1_equals_mac_pytorch": true, "3_coreml_fp32_orders_equal_mac_pytorch_all_methods": true, "3_kv_fp32_top1_equals_mac_pytorch": true, "3_kv_fp32_orders_equal_mac_pytorch_all_methods": true, "4_int8_strict_logits_alignment_passed (expected false)": false, "4_int8_kv_vs_same_weight_stateless_int8_passed": true, "4_kv_int8_top1_equals_stateless_int8": true, "4_kv_int8_orders_equal_stateless_int8_all_methods": false, "7_kvchunk_fp32_alignment_passed": true, "7_kvchunk_int8_vs_same_weight_stateless_int8_passed": true, "7_kvchunk_int8_strict_vs_fp32_fixtures_passed (expected false)": false, "7_kvchunk_fp32_top1_equals_stateless_fp32": true, "7_kvchunk_fp32_orders_equal_stateless_fp32_all_methods": true, "7_kvchunk_int8_top1_equals_stateless_int8": true, "7_kvchunk_int8_orders_equal_stateless_int8_all_methods": false, "8_v21_coreml_fp32_top1_equals_recorded": true, "5_measurements_recorded": true, "6_ios_resources_prepared": true }, "int8_vs_coreml_fp32": { "standard-development": { "fp32": 523, "int8": 522, "v1_wrong_v2_correct": 2, "v1_correct_v2_wrong": 3, "both_correct": 520, "both_wrong": 203, "net_correct": -1, "exact_mcnemar_p_two_sided": 1, "significant_at_0_05": false, "order_changed_cases": 254, "top1_changed_cases": 11, "max_score_sum_abs": 0.7504582202527672, "mean_score_sum_abs": 0.12163806840569316 }, "legacy-regression": { "fp32": 370, "int8": 371, "v1_wrong_v2_correct": 1, "v1_correct_v2_wrong": 0, "both_correct": 370, "both_wrong": 129, "net_correct": 1, "exact_mcnemar_p_two_sided": 1, "significant_at_0_05": false, "order_changed_cases": 46, "top1_changed_cases": 2, "max_score_sum_abs": 0.8114978885278106, "mean_score_sum_abs": 0.11693049636118634 }, "ajimee": { "fp32": 139, "int8": 139, "v1_wrong_v2_correct": 0, "v1_correct_v2_wrong": 0, "both_correct": 139, "both_wrong": 61, "net_correct": 0, "exact_mcnemar_p_two_sided": 1.0, "significant_at_0_05": false, "order_changed_cases": 106, "top1_changed_cases": 0, "max_score_sum_abs": 0.9850609144195914, "mean_score_sum_abs": 0.1471606100397155 }, "historical-development": { "fp32": 123, "int8": 123, "v1_wrong_v2_correct": 0, "v1_correct_v2_wrong": 0, "both_correct": 123, "both_wrong": 14, "net_correct": 0, "exact_mcnemar_p_two_sided": 1.0, "significant_at_0_05": false, "order_changed_cases": 88, "top1_changed_cases": 0, "max_score_sum_abs": 0.7235050201416016, "mean_score_sum_abs": 0.09826715727665117 } }, "v21_int8_vs_v42_int8": { "standard-development": { "v1_wrong_v2_correct": 44, "v1_correct_v2_wrong": 28, "both_correct": 478, "both_wrong": 178, "net_correct": 16, "exact_mcnemar_p_two_sided": 0.07636915467183139 }, "legacy-regression": { "v1_wrong_v2_correct": 13, "v1_correct_v2_wrong": 17, "both_correct": 358, "both_wrong": 112, "net_correct": -4, "exact_mcnemar_p_two_sided": 0.584664711728692 }, "ajimee": { "v1_wrong_v2_correct": 9, "v1_correct_v2_wrong": 14, "both_correct": 130, "both_wrong": 47, "net_correct": -5, "exact_mcnemar_p_two_sided": 0.4048728942871094 }, "historical-development": { "v1_wrong_v2_correct": 4, "v1_correct_v2_wrong": 3, "both_correct": 119, "both_wrong": 11, "net_correct": 1, "exact_mcnemar_p_two_sided": 1 } }, "alignment": { "v42-fp32-alignment-v1": { "passed": true, "checks": 5, "max_abs": 5.245208740234375e-05, "failed_checks": 0 }, "v42-kv-fp32-alignment-v1": { "passed": true, "checks": 17, "max_abs": 4.9591064453125e-05, "failed_checks": 0 }, "v42-int8-alignment-v1": { "passed": false, "checks": 5, "max_abs": 0.4774494171142578, "failed_checks": 5 }, "v42-kv-int8-vs-stateless-int8-v1": { "passed": true, "checks": 16, "max_abs": 4.00543212890625e-05, "failed_checks": 0 }, "v42-kv-int8-vs-fp32-fixtures-v1": { "passed": false, "checks": 17, "max_abs": 0.707888126373291, "failed_checks": 17 }, "v42-kvchunk-fp32-alignment-v1": { "passed": true, "checks": 27, "max_abs": 5.245208740234375e-05, "failed_checks": 0 }, "v42-kvchunk-int8-vs-stateless-int8-v1": { "passed": true, "checks": 26, "max_abs": 4.00543212890625e-05, "failed_checks": 0 }, "v42-kvchunk-int8-vs-fp32-fixtures-v1": { "passed": false, "checks": 22, "max_abs": 0.7078914642333984, "failed_checks": 22 } } }, "notation_flip_analysis_vs_v21": { "standard-development": { "A": { "improved": 28, "regressed": 23, "net": 5 }, "S": { "improved": 15, "regressed": 5, "net": 10 }, "P": { "improved": 3, "regressed": 0, "net": 3 }, "meaning_plus_proper_exact_p": 0.010622024536132812, "flips_margin_lt_1nat": { "all": 19, "of": 74, "A": 16 } }, "legacy-regression": { "A": { "improved": 12, "regressed": 17, "net": -5 }, "S": { "improved": 3, "regressed": 2, "net": 1 }, "P": { "improved": 0, "regressed": 0, "net": 0 }, "meaning_plus_proper_exact_p": 1.0, "flips_margin_lt_1nat": { "all": 12, "of": 34, "A": 11 } }, "ajimee": { "A": { "improved": 1, "regressed": 2, "net": -1 }, "S": { "improved": 7, "regressed": 11, "net": -4 }, "P": { "improved": 0, "regressed": 0, "net": 0 }, "meaning_plus_proper_exact_p": 0.480682373046875, "flips_margin_lt_1nat": { "all": 6, "of": 21, "A": 2 } }, "historical-development": { "A": { "improved": 3, "regressed": 1, "net": 2 }, "S": { "improved": 1, "regressed": 2, "net": -1 }, "P": { "improved": 0, "regressed": 0, "net": 0 }, "meaning_plus_proper_exact_p": 1.0, "flips_margin_lt_1nat": { "all": 3, "of": 7, "A": 2 } }, "reference725": { "A": { "improved": 27, "regressed": 24, "net": 3 }, "S": { "improved": 11, "regressed": 7, "net": 4 }, "P": { "improved": 2, "regressed": 3, "net": -1 }, "meaning_plus_proper_exact_p": 0.6776394844055176, "flips_margin_lt_1nat": { "all": 23, "of": 74, "A": 19 } } }, "measurements": { "mac": { "source_sha256": "66eeb948028365445bb1a8dad6e84e9f8d69806ac9702dd8293821b1a47e2a53", "hardware": "Apple M3, macOS 27.2, CPU_ONLY", "stateless_int8_warm_p50_ms": { "T1": 1.3, "T128": 6.07 }, "v21_int8_warm_p50_ms": { "T1": 0.99, "T128": 5.68 }, "kvchunk_int8_prefill_p50_ms": { "P7": 2.49, "P31": 3.52, "P63": 4.55, "P119": 6.56 }, "v21_kv_int8_prefill_p50_ms": { "P7": 2.19, "P31": 3.03, "P63": 4.13, "P119": 6.11 } }, "iphone16promax_xctest_host": { "source_sha256": "3c8566cef8ad78b11d5a9d1bf3eb55d6eadef8aaf3c1d0f5aa7680615a562cbb", "conditions": "iOS 27.2, wired and charging, thermal state fair", "v42_int8": { "predict_T128_p50_ms": 4.9, "rerank_p50_ms": 25.1, "rerank_p95_ms": 120.9, "beam_p50_ms": 108.9, "next_words_p50_ms": 16.1, "footprint_peak_mb": 44.35 }, "v21_int8": { "predict_T128_p50_ms": 4.55, "rerank_p50_ms": 21.3, "rerank_p95_ms": 136.8, "beam_p50_ms": 101.8, "next_words_p50_ms": 13.2, "footprint_peak_mb": 49.95 }, "scope": "XCTest host process, not the keyboard extension" } }, "deployment_status": "Research release. V2.1 remains the deployed default in Vime; V4.2 is an opt-in variant.", "source_reports_sha256": { "outputs/training-v42-20261020/remote/records/summary.json": "b15397eab9a11d732fe8465d2eecd17a840e25dad15714b6d6c35f59aba2d4ac", "outputs/evaluation-v42-20261020/standard-development/metrics.json": "17a03af37beaa494ab307bfa8199867dac5da8b601d416f500e73865abf23a28", "outputs/evaluation-v42-20261020/legacy-regression/metrics.json": "bbadc895779a62bd3b897efa8b8fcffa18ccd9563cb3d68b1211c7f636ef4864", "outputs/evaluation-v42-20261020/ajimee/metrics.json": "4c90bdea27b3876b58714eec204c5bef3a74c8030852bb8b622585d05eeab83d", "outputs/evaluation-v42-20261020/historical-development/metrics.json": "84d5e4008dcd372868c91713411be3bc3981a1f232f859e327bd32aaa32129e1", "outputs/evaluation-v42-20261020/plan.json": "d47df6154c193088b865782a1048794d542c2ecc2ffec938a3cd1a5eac2e579d", "outputs/evaluation-v42-reference-20261020/reference725/metrics.json": "5d7b2db2601317f68c7876f34ad19c41b17ca0c1ed0509cc419ceedd27dade3a", "outputs/evaluation-v42-reference-20261020/reference-test-plan.json": "76105753a0e65b08f59832b80db10ea549cf9a1c987ea3b98b83241b4bfd8de5", "outputs/evaluation-v42-reference-20261020/paired-v21.json": "0f824aa5a6513409d82655ff46e8a71750629d658311f45acc06dbecd5a0bf1f", "outputs/deployment/v42-mac-acceptance-v2/status.json": "f27981649fbaf85fffa99ce4d7f5a0337c6dd2c6fddfaa20da09e6e7c62f31df", "outputs/analysis-v42-regressions-20261020/summary.json": "c7d2e2d6c6edc24a3765d9761a11bca23d7a763ddde56b30ddbaf6cec4cb3ade" }, "scope": "Aggregate metrics only. No corpus, benchmark text, candidate pools, per-case scores, case identifiers, user inputs or raw device traces are distributed." }