Voltline's picture
Release VimeML V2.1 step40000 FP32 and Core ML INT8 (GPL-2.0)
29f25be verified
Raw History Blame Contribute Delete
14.5 kB
{
"format": "vimeml_public_evaluation_v2",
"model_version": "2.1-extend5-step40000-int8-b32-v1",
"checkpoint_sha256": "fa202f18d00cf5e77469bd3be5b8b264117fb73233ec9966fa3213257a470104",
"coreml_manifest_sha256": "1dda48c76eacfa3ba99764098fab721eed4ada413130b99c0324620365aca0b7",
"training": {
"full_validation_bpc_fp32": 3.0802686334,
"test_used_for_selection": false,
"v20_epochs": 4,
"restart1_epochs": 2,
"extend5_stopped_step": 161095,
"selected_extend5_step": 40000
},
"ranking": {
"ajimee": {
"original_order": {
"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
},
"fp32_lm": {
"cases": 200,
"top1_correct": 144,
"top1_accuracy": 0.72,
"top5_correct": 160,
"top5_accuracy": 0.8,
"candidate_pool_covered": 162,
"candidate_pool_recall": 0.81,
"covered_top1_accuracy": 0.8888888888888888,
"mean_min_cer": 0.030729146273316313,
"mrr": 0.7542142857142857,
"corrected_vs_azookey": 61,
"regressed_vs_azookey": 4,
"fallback_cases": 0
},
"int8_lm": {
"cases": 200,
"top1_correct": 144,
"top1_accuracy": 0.72,
"top5_correct": 160,
"top5_accuracy": 0.8,
"candidate_pool_covered": 162,
"candidate_pool_recall": 0.81,
"covered_top1_accuracy": 0.8888888888888888,
"mean_min_cer": 0.03116485517226051,
"mrr": 0.7533809523809524,
"corrected_vs_azookey": 62,
"regressed_vs_azookey": 5,
"fallback_cases": 0
},
"fp32_comparison": {
"order_changed_cases": 163,
"top1_changed_cases": 7,
"max_score_sum_abs": 1.8200132250785828,
"mean_score_sum_abs": 0.2882139610526506
},
"labels_formal_gold": null
},
"development": {
"original_order": {
"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
},
"fp32_lm": {
"cases": 137,
"top1_correct": 122,
"top1_accuracy": 0.8905109489051095,
"top5_correct": 136,
"top5_accuracy": 0.9927007299270073,
"candidate_pool_covered": 136,
"candidate_pool_recall": 0.9927007299270073,
"covered_top1_accuracy": 0.8970588235294118,
"mean_min_cer": 0.029006070976873896,
"mrr": 0.936374695863747,
"corrected_vs_azookey": 14,
"regressed_vs_azookey": 3,
"fallback_cases": 0
},
"int8_lm": {
"cases": 137,
"top1_correct": 122,
"top1_accuracy": 0.8905109489051095,
"top5_correct": 136,
"top5_accuracy": 0.9927007299270073,
"candidate_pool_covered": 136,
"candidate_pool_recall": 0.9927007299270073,
"covered_top1_accuracy": 0.8970588235294118,
"mean_min_cer": 0.029006070976873896,
"mrr": 0.9351581508515816,
"corrected_vs_azookey": 14,
"regressed_vs_azookey": 3,
"fallback_cases": 0
},
"fp32_comparison": {
"order_changed_cases": 121,
"top1_changed_cases": 1,
"max_score_sum_abs": 1.3734395503997803,
"mean_score_sum_abs": 0.19201541958957397
},
"labels_formal_gold": null
},
"expanded-development": {
"original_order": {
"cases": 2000,
"top1_correct": 1224,
"top1_accuracy": 0.612,
"top5_correct": 1711,
"top5_accuracy": 0.8555,
"candidate_pool_covered": 1756,
"candidate_pool_recall": 0.878,
"covered_top1_accuracy": 0.6970387243735763,
"mean_min_cer": 0.049437279587549655,
"mrr": 0.714617032967033,
"corrected_vs_azookey": 0,
"regressed_vs_azookey": 0,
"fallback_cases": 0
},
"fp32_lm": {
"cases": 2000,
"top1_correct": 1487,
"top1_accuracy": 0.7435,
"top5_correct": 1748,
"top5_accuracy": 0.874,
"candidate_pool_covered": 1756,
"candidate_pool_recall": 0.878,
"covered_top1_accuracy": 0.8468109339407744,
"mean_min_cer": 0.029395604106192606,
"mrr": 0.8023303571428563,
"corrected_vs_azookey": 322,
"regressed_vs_azookey": 59,
"fallback_cases": 0
},
"int8_lm": {
"cases": 2000,
"top1_correct": 1481,
"top1_accuracy": 0.7405,
"top5_correct": 1748,
"top5_accuracy": 0.874,
"candidate_pool_covered": 1756,
"candidate_pool_recall": 0.878,
"covered_top1_accuracy": 0.8433940774487472,
"mean_min_cer": 0.02966218634430054,
"mrr": 0.800903769841269,
"corrected_vs_azookey": 319,
"regressed_vs_azookey": 62,
"fallback_cases": 0
},
"fp32_comparison": {
"order_changed_cases": 404,
"top1_changed_cases": 30,
"max_score_sum_abs": 1.790113139897585,
"mean_score_sum_abs": 0.22886438840285703
},
"labels_formal_gold": false
}
},
"alignment": {
"fp32": {
"strict_passed": true,
"tolerances": {
"atol": 0.0003,
"rtol": 0.0003
},
"logits": [
{
"max_abs": 9.5367431640625e-06,
"mean_abs": 1.2576065273606218e-06,
"rmse": 1.6289324748934612e-06,
"outside_tolerance": 0,
"values": 16384,
"id": "bos",
"length": 1,
"valid_prefix": 1,
"last_top1_equal": true
},
{
"max_abs": 3.910064697265625e-05,
"mean_abs": 3.185306411523925e-06,
"rmse": 4.141091950346693e-06,
"outside_tolerance": 0,
"values": 262144,
"id": "short",
"length": 16,
"valid_prefix": 16,
"last_top1_equal": true
},
{
"max_abs": 4.57763671875e-05,
"mean_abs": 4.321694975506236e-06,
"rmse": 5.538972619049361e-06,
"outside_tolerance": 0,
"values": 2097152,
"id": "long",
"length": 128,
"valid_prefix": 128,
"last_top1_equal": true
},
{
"max_abs": 4.482269287109375e-05,
"mean_abs": 3.3942287953614425e-06,
"rmse": 4.3666045771710725e-06,
"outside_tolerance": 0,
"values": 262144,
"id": "right_pad",
"length": 128,
"valid_prefix": 16,
"last_top1_equal": true
},
{
"max_abs": 4.482269287109375e-05,
"mean_abs": 3.3942287953614425e-06,
"rmse": 4.3666045771710725e-06,
"outside_tolerance": 0,
"values": 262144,
"id": "changed_future",
"length": 128,
"valid_prefix": 16,
"last_top1_equal": true
}
],
"invariants": {
"right_pad": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rmse": 0.0,
"outside_tolerance": 0,
"values": 262144
},
"changed_future": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rmse": 0.0,
"outside_tolerance": 0,
"values": 262144
}
}
},
"int8": {
"strict_passed": false,
"tolerances": {
"atol": 0.0003,
"rtol": 0.0003
},
"logits": [
{
"max_abs": 0.1717967987060547,
"mean_abs": 0.03132702315087954,
"rmse": 0.03977237435193969,
"outside_tolerance": 15692,
"values": 16384,
"id": "bos",
"length": 1,
"valid_prefix": 1,
"last_top1_equal": true
},
{
"max_abs": 0.8956146240234375,
"mean_abs": 0.10476142495688912,
"rmse": 0.13775674684387576,
"outside_tolerance": 257685,
"values": 262144,
"id": "short",
"length": 16,
"valid_prefix": 16,
"last_top1_equal": true
},
{
"max_abs": 1.0094914436340332,
"mean_abs": 0.10851713189305306,
"rmse": 0.1394378905311297,
"outside_tolerance": 2067332,
"values": 2097152,
"id": "long",
"length": 128,
"valid_prefix": 128,
"last_top1_equal": true
},
{
"max_abs": 0.8956174850463867,
"mean_abs": 0.10476151736959416,
"rmse": 0.13775683790182477,
"outside_tolerance": 257684,
"values": 262144,
"id": "right_pad",
"length": 128,
"valid_prefix": 16,
"last_top1_equal": true
},
{
"max_abs": 0.8956174850463867,
"mean_abs": 0.10476151736959416,
"rmse": 0.13775683790182477,
"outside_tolerance": 257684,
"values": 262144,
"id": "changed_future",
"length": 128,
"valid_prefix": 16,
"last_top1_equal": true
}
],
"invariants": {
"right_pad": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rmse": 0.0,
"outside_tolerance": 0,
"values": 262144
},
"changed_future": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rmse": 0.0,
"outside_tolerance": 0,
"values": 262144
}
}
}
},
"fixed_generation": {
"next_token_top1_equal": 12,
"greedy_32_token_equal": 8,
"total": 12
},
"quantization_diagnostics": {
"format": "vimeml_quantization_analysis_v1",
"date": "2026-10-08",
"source_paths": [
"outputs/ime-eval/expanded-v21-dev-draft-extend5-best/scores.jsonl",
"outputs/deployment/v21-int8-expanded-development-v1/scores.jsonl"
],
"scope": "Existing frozen draft development scores; no model execution, blind data or label edits",
"cases": 2000,
"top1_changes": 30,
"wins": 8,
"losses": 14,
"paired_exact_two_sided_p": 0.28627872467041016,
"fp32_top1_margin_by_change": {
"changed": {
"cases": 30,
"median_logp_sum": 0.07003979850560427,
"maximum_logp_sum": 0.36277501191943884,
"count_margin_le_0_25": 27
},
"unchanged": {
"cases": 1910,
"median_logp_sum": 3.0718008121475577,
"maximum_logp_sum": 26.55984030291438,
"count_margin_le_0_25": 85
}
},
"interpretation": "No statistically significant accuracy decrease detected; this does not establish equivalence or formal gold quality."
},
"physical_keyboard_extension": {
"device": "iPhone 16 Pro Max",
"os": "iOS 27.2 (24B5099f)",
"pid": 17014,
"model_version": "2.1-extend5-step40000-int8-b32-v1",
"instruments_duration_seconds": 201.29728,
"instruments_samples": 164,
"in_process_duration_seconds": 191.90217375,
"in_process_samples": 1920,
"kernel_lifetime_footprint_peak_mib": 34.720596,
"in_process_sampled_footprint_peak_mib": 28.298721,
"instruments_sampled_footprint_peak_mib": 28.001846,
"instruments_private_resident_peak_mib": 73.78125,
"lm_scoring_p95_ms": 53.915625,
"lm_next_words_p95_ms": 38.114333,
"measurement_enabled_after_completion": false,
"input_text_recorded": false,
"scope": "Bounded real extension workload; separate from test host, sampled resident values and lifetime physical footprint"
},
"limitations": [
{
"item": "INT8 strict frozen logits alignment",
"result": "failed",
"max_absolute_error": 1.0094914436340332,
"atol": 0.0003,
"rtol": 0.0003
},
{
"item": "Expanded development draft Top-1",
"fp32": 1487,
"int8": 1481,
"total": 2000,
"delta_percentage_points": -0.3
},
{
"item": "Actual extension result-to-UI long tail",
"maximum_ms": 1926.0947,
"cause": "unlocated; no per-event correlation timestamps"
},
{
"item": "Late-window extension footprint growth and second model load",
"cause": "unlocated; insufficient evidence for leak/cache/host attribution"
},
{
"item": "Fully cold startup",
"status": "not certified: measured first process load may use warm OS/Core ML caches"
},
{
"item": "Long-term footprint plateau/system termination audit",
"status": "not certified by this bounded workload"
}
],
"experiment_decision": "V2 series complete; current INT8 task-quality loss and observed UI tail accepted.",
"source_reports_sha256": {
"outputs/deployment/v21-int8-ajimee-v1/metrics.json": "7b136dbed85b2f354eeae1838eb7f32e34c8eb3e909fe4fcce1ac449f1480c63",
"outputs/deployment/v21-int8-development-v1/metrics.json": "f1384fa5d5b7c4da2503a9118300a0857bf2ef5c4ee2d7213817851839f2394c",
"outputs/deployment/v21-int8-expanded-development-v1/metrics.json": "0a07dbe706add83b3c3e917dce1334d0227f105d29347671bca539442f887226",
"outputs/deployment/v21-fp32-alignment-v1/alignment.json": "3c0a368f4836388733fcb21882e431b8b7548b1e9ea61fefeeb479ce51954b0b",
"outputs/deployment/v21-int8-alignment-v1/alignment.json": "791d047cf14d15025a33bc51c0e3e1ebaa986247832bdd42fcda027f8901db7f",
"outputs/deployment/v21-closeout-v1/acceptance.json": "1992372e8867f29c0d2add7d9d15ab0ace302555e42741c821bad8d6c0e20e3d",
"outputs/deployment/v21-quantization-analysis-v1/analysis.json": "0fd04daf4e8b86a6cb028d89f4c598312bdafcbca271f2b27d445b6db581a132"
},
"scope": "Saved fixed-pool metrics and bounded device workload; draft expanded labels, no blind scoring. No benchmark texts, per-case scores, user inputs or raw device traces distributed."
}