tyger / evaluation /metrics.json
suryatmodulus's picture Simonlob's picture
Duplicate from nineninesix/tyger
7800076
Raw History Blame Contribute Delete
10.6 kB
{
"schema_version": 1,
"model": "nineninesix/tyger-s2s",
"language": "en",
"evaluation_type": "offline_controller_replay",
"evaluation_policy": {
"ep_thr": 0.91,
"ep_deb": 2,
"barge_thr": 0.93,
"barge_deb": 2,
"bc_thr": 0.8,
"bc_deb": 2,
"bc_refractory_frames": 15,
"fallback_sil_frames": 15,
"vad_db": -50.0,
"min_speech_ms": 350.0,
"warmup_frames": 6
},
"default_policy_evaluated": false,
"datasets": {
"open_yap_1k": {
"name": "Open Yap 1K",
"version": "1.0",
"url": "https://theagenticdatacompany.com/open-yap-1k",
"train_conversations": 1395,
"dev_conversations": 107,
"test_conversations": 100,
"split_unit": "conversation",
"channel_assignments_grouped": true,
"test_frame_windows": 23936
},
"full_duplex_bench": {
"name": "Full-Duplex-Bench",
"version": "1.0",
"url": "https://huggingface.co/datasets/Ssshangfu/Full-Duplex-Bench-Data",
"revision": "a3579cb9ebbe164204b196e33d4d0e91860eb90e",
"subsets": [
"candor_turn_taking",
"candor_pause_handling"
],
"scope": "Adapted endpoint replay; not the official full voice-agent benchmark or TOR."
},
"internal": {
"name": "Internal evaluation",
"access": "non-public",
"scope": "Additional out-of-domain endpoint probes; source identifiers are not distributed."
}
},
"open_yap_test_frames": {
"yield_auc": 0.9796688041466441,
"yield_ap": 0.6657115340151823,
"yield_positive_frames": 78882,
"yield_masked_frames": 2522031,
"bc_auc_freerun": 0.8975721507406702,
"bc_ap": 0.30040181934186144,
"bc_positive_frames": 31352,
"bc_masked_frames": 3030374,
"ep_auc_freerun": 0.9838886216113089,
"endpoint_ap": 0.8149145794136884,
"endpoint_positive_frames": 163373,
"endpoint_masked_frames": 2651747,
"state_macro_f1": 0.9436201457545479,
"state_confusion": [
[
2987221,
9814,
17284,
5521
],
[
7876,
314958,
24061,
34
],
[
17528,
27440,
2464707,
1748
],
[
6912,
11,
1385,
71910
]
],
"hazard_accuracy": 0.9315253565968102
},
"open_yap_test_events": {
"yield_detect": 0.2531545741324921,
"yield_latency_ms": 400.0,
"yield_false_per_h": 0.0,
"bc_recall": 0.14856081708449395,
"bc_latency_ms": -160.0,
"bc_acks_per_h": 135.0786866135613,
"yield_onsets": 1268,
"backchannel_onsets": 1077,
"clean_speech_hours": 4.722822222222222,
"listen_hours_approx": 6.86,
"conversation_prefix_seconds": 240
},
"endpoints": {
"open_yap_dev": {
"fallback_on": {
"mean_latency_s": 0.4745854700854754,
"median_latency_s": 0.33249999999994584,
"p90_latency_s": 0.9725000000001192,
"head_share": 0.8316666666666667,
"fallback_share": 0.14333333333333334,
"no_endpoint_share": 0.025,
"miss_within_1p2s": 0.058333333333333334,
"false_cutins_per_floor_h": 4.498953731690305,
"turns_with_early_cutin_share": 0.285,
"n_turn": 600,
"n_pause": 600,
"deadline_eligible_turns": 600,
"deadline_censored_turns": 0,
"floor_h": 57.346666666666664
},
"head_only": {
"mean_latency_s": 0.4530970149253823,
"median_latency_s": 0.2525000000001185,
"p90_latency_s": 0.8924999999999983,
"head_share": 0.8933333333333333,
"fallback_share": 0.0,
"no_endpoint_share": 0.10666666666666667,
"miss_within_1p2s": 0.16166666666666665,
"false_cutins_per_floor_h": 3.487561032318066,
"turns_with_early_cutin_share": 0.25833333333333336,
"n_turn": 600,
"n_pause": 600,
"deadline_eligible_turns": 600,
"deadline_censored_turns": 0,
"floor_h": 57.346666666666664
}
},
"open_yap_test": {
"fallback_on": {
"mean_latency_s": 0.46497668393782077,
"median_latency_s": 0.2525000000001185,
"p90_latency_s": 0.9725000000001908,
"head_share": 0.843,
"fallback_share": 0.122,
"no_endpoint_share": 0.035,
"miss_within_1p2s": 0.074,
"false_cutins_per_floor_h": 6.537794226996417,
"turns_with_early_cutin_share": 0.299,
"n_turn": 1000,
"n_pause": 1000,
"deadline_eligible_turns": 1000,
"deadline_censored_turns": 0,
"floor_h": 63.477066666666666
},
"head_only": {
"mean_latency_s": 0.4364999999999966,
"median_latency_s": 0.25250000000003947,
"p90_latency_s": 0.972499999999957,
"head_share": 0.9,
"fallback_share": 0.0,
"no_endpoint_share": 0.1,
"miss_within_1p2s": 0.16,
"false_cutins_per_floor_h": 5.009683287192435,
"turns_with_early_cutin_share": 0.275,
"n_turn": 1000,
"n_pause": 1000,
"deadline_eligible_turns": 1000,
"deadline_censored_turns": 0,
"floor_h": 63.477066666666666
}
},
"internal": {
"fallback_on": {
"mean_latency_s": 0.2865495867768678,
"median_latency_s": 0.2525000000000013,
"p90_latency_s": 0.4924999999999995,
"head_share": 0.9834710743801653,
"fallback_share": 0.01652892561983471,
"no_endpoint_share": 0.0,
"miss_within_1p2s": 0.0,
"false_cutins_per_floor_h": 28.316057268126336,
"turns_with_early_cutin_share": 0.12396694214876033,
"n_turn": 242,
"n_pause": 539,
"deadline_eligible_turns": 242,
"deadline_censored_turns": 0,
"floor_h": 6.568711111111111
},
"head_only": {
"mean_latency_s": 0.2893464730290536,
"median_latency_s": 0.2525000000000013,
"p90_latency_s": 0.4924999999999975,
"head_share": 0.9958677685950413,
"fallback_share": 0.0,
"no_endpoint_share": 0.004132231404958678,
"miss_within_1p2s": 0.012396694214876033,
"false_cutins_per_floor_h": 25.575793661533467,
"turns_with_early_cutin_share": 0.12396694214876033,
"n_turn": 242,
"n_pause": 539,
"deadline_eligible_turns": 242,
"deadline_censored_turns": 0,
"floor_h": 6.568711111111111
}
},
"fdb_turn_taking": {
"fallback_on": {
"mean_latency_s": 0.3830172413793251,
"median_latency_s": 0.3074999999999104,
"p90_latency_s": 0.6324999999999386,
"head_share": 0.9243697478991597,
"fallback_share": 0.05042016806722689,
"no_endpoint_share": 0.025210084033613446,
"miss_within_1p2s": 0.05042016806722689,
"false_cutins_per_floor_h": 0.0,
"turns_with_early_cutin_share": 0.25210084033613445,
"n_turn": 119,
"n_pause": 0,
"deadline_eligible_turns": 119,
"deadline_censored_turns": 0,
"floor_h": 0.4636100868055557
},
"head_only": {
"mean_latency_s": 0.4245175438596653,
"median_latency_s": 0.30249999999994737,
"p90_latency_s": 0.602500000000009,
"head_share": 0.957983193277311,
"fallback_share": 0.0,
"no_endpoint_share": 0.04201680672268908,
"miss_within_1p2s": 0.07563025210084033,
"false_cutins_per_floor_h": 0.0,
"turns_with_early_cutin_share": 0.24369747899159663,
"n_turn": 119,
"n_pause": 0,
"deadline_eligible_turns": 119,
"deadline_censored_turns": 0,
"floor_h": 0.4636100868055557
}
},
"fdb_pause_handling": {
"fallback_on": {
"mean_latency_s": null,
"median_latency_s": null,
"p90_latency_s": null,
"head_share": null,
"fallback_share": null,
"no_endpoint_share": null,
"miss_within_1p2s": null,
"false_cutins_per_floor_h": 173.25028170777514,
"turns_with_early_cutin_share": null,
"n_turn": 0,
"n_pause": 277,
"deadline_eligible_turns": 0,
"deadline_censored_turns": 0,
"floor_h": 0.7099555555555556
},
"head_only": {
"mean_latency_s": null,
"median_latency_s": null,
"p90_latency_s": null,
"head_share": null,
"fallback_share": null,
"no_endpoint_share": null,
"miss_within_1p2s": null,
"false_cutins_per_floor_h": 147.8965819456617,
"turns_with_early_cutin_share": null,
"n_turn": 0,
"n_pause": 277,
"deadline_eligible_turns": 0,
"deadline_censored_turns": 0,
"floor_h": 0.7099555555555556
}
}
},
"open_yap_dev_filler_proxy": {
"threshold": 0.91,
"recall": 0.6199889301810706,
"n_turn_ends": 12647,
"median_latency_ms": 80.0,
"false_endpoints_per_floor_h": 1004.5060622336916,
"filler_floor_takings": 390,
"user_floor_h": 57.38442222222223
},
"controller_compute": {
"device": "NVIDIA GeForce RTX 4090",
"batch_size": 1,
"prefill_frames": 64,
"measured_frames_per_repeat": 200,
"repeats": 3,
"parameter_count": 13991432,
"parameter_mib": 53.373077392578125,
"peak_allocated_mib": 108.75830078125,
"wall_ms": {
"p50": 0.805371964816004,
"p90": 0.8225860772654414
},
"cuda_event_ms": {
"p50": 0.7813119888305664,
"p90": 0.7976959943771362
},
"streaming_state_bytes_batch1": 9216,
"precision": "fp32",
"tf32": false,
"includes_encoder": false,
"includes_asr": false,
"includes_tts": false,
"includes_audio_wait": false
},
"limitations": [
"One training seed; evaluation sets informed development and are diagnostic.",
"State and yield frame metrics use teacher-forced previous state; endpoint and backchannel use forced LISTEN.",
"Event metrics use frozen development-set thresholds, not the lower deployment yield/backchannel defaults.",
"Endpoint latency is conditional on firing; report misses and early cut-ins alongside latency.",
"Full-Duplex-Bench results cover adapted v1.0 CANDOR probes only, not FDB-v3 or official full-agent scores.",
"Internal evaluation data are non-public, so that subset is not independently reproducible from this release.",
"Controller compute excludes Nemotron, ASR, audio buffering and all downstream voice-agent components.",
"This release does not add a new live conversation benchmark or a new ASR quality benchmark."
]
}