Release FlashVAD v0.1 alpha research preview

MODEL_LICENSE.md

@@ -0,0 +1,21 @@
+# FlashVAD v0.1 model licence
+
+The model artifacts in this repository are available under the
+[Creative Commons Attribution 4.0 International licence](https://creativecommons.org/licenses/by/4.0/).
+
+This applies to:
+
+- `flashvad-v0.1.pt`;
+- `flashvad-stream.onnx`;
+- the accompanying model configuration and calibration metadata;
+- generated native weight arrays such as `native/macos/flashvad_weights.c`
+  and `native/macos/flashvad_weights.h`, including their packaged-wheel paths.
+
+Attribution:
+
+> FlashVAD v0.1 by OSS Codes, with training-data and teacher attribution
+> documented in the repository `NOTICE` file.
+
+The repository source code, excluding these model-derived artifacts, is
+separately licensed under the MIT licence.
+Third-party datasets, models, and benchmark materials retain their own terms.

NOTICE

@@ -0,0 +1,23 @@
+FlashVAD
+Copyright 2026 OSS Codes
+
+The FlashVAD v0.1 research checkpoint was trained using derived data from:
+
+- FLEURS, Google LLC, licensed under CC BY 4.0.
+  https://huggingface.co/datasets/google/fleurs
+- AMI Meeting Corpus, licensed under CC BY 4.0.
+  https://groups.inf.ed.ac.uk/ami/corpus/
+- MUSAN, David Snyder, Guoguo Chen, and Daniel Povey, licensed under CC BY 4.0.
+  https://www.openslr.org/17/
+
+Weak frame targets were generated with Silero VAD, which is distributed under
+the MIT License:
+https://github.com/snakers4/silero-vad
+
+TEN VAD's public test set was used only for the retained cross-domain
+evaluation summary. TEN source code and weights are not included:
+https://github.com/TEN-framework/ten-vad
+
+FireRedVAD source code and weights are not included. Its official release was
+used only for reference measurements and rejected teacher-student experiments:
+https://github.com/FireRedTeam/FireRedVAD

README.md

@@ -0,0 +1,159 @@
+---
+license: cc-by-4.0
+library_name: flashvad
+language:
+  - ar
+  - en
+  - gu
+  - hi
+  - kn
+  - pa
+  - ta
+  - te
+  - ur
+tags:
+  - audio
+  - voice-activity-detection
+  - vad
+  - streaming
+  - onnx
+  - pytorch
+  - telephony
+  - webrtc
+---
+
+# FlashVAD v0.1 model card
+
+## Summary
+
+FlashVAD v0.1 is a 46,170-parameter causal streaming voice-activity detector.
+It consumes 16 kHz mono audio, produces one speech probability every 10 ms,
+and keeps independent convolutional, recurrent, feature, and detector state
+per call.
+
+This checkpoint is an **alpha research preview** for integration testing,
+browser demonstrations, and shadow evaluation. It is not approved for
+production use and is not validated as a general multilingual or India/GCC
+call model.
+
+## Artifacts
+
+| Artifact | SHA-256 |
+|---|---|
+| `flashvad-v0.1.pt` | `ca9e35475518466b2a1f2e89b4953cd1e26e3d8c513cdcf265ab319e74e2b288` |
+| `flashvad-stream.onnx` | `9a88e34bf3118d60e25a16cb622cb394e2f3ab71445b0aa5957df6f1d5f1b6ba` |
+| `config.json` | `0b1ad372808f7c67cea5a1ca4b41a817714c9a0b0cf49b3baa56fe8d5f64ad2b` |
+| `detector-calibration.json` | `b5d000e0406d81fbd87a9e66194a877fa8433a76783faac8275c7969c43051b4` |
+
+The ONNX graph accepts precomputed 43-dimensional causal features. Use
+`src/flashvad/features.py` or
+`report-site/src/lib/vad-features.mjs`; it is not a raw-waveform graph.
+
+The public ONNX file is self-contained and stripped of exporter stack traces,
+local paths, and private build metadata.
+
+## Download and source
+
+Download the complete model repository:
+
+```bash
+hf download oss-codes/flashvad --local-dir flashvad-model
+```
+
+Source code and runtime integrations are published separately at
+[`oss-codes/flashvad`](https://github.com/oss-codes/flashvad).
+
+The ONNX graph does not accept raw waveform audio. It expects the causal
+43-dimensional features described below, with independent feature and model
+state for every call.
+
+## Intended use
+
+Appropriate current uses:
+
+- research and architecture evaluation;
+- functional integration with browser, LiveKit, Pipecat, SIP, or PSTN stacks;
+- latency and concurrency measurement;
+- shadow-mode comparison on consented, labelled call audio.
+
+Do not use this checkpoint as the sole basis for:
+
+- emergency, medical, legal, financial, or safety-critical decisions;
+- call recording consent or compliance decisions;
+- a production multilingual accuracy claim;
+- semantic end-of-turn detection.
+
+Reset all feature, model, resampler, and detector state when a call ends or an
+audio discontinuity occurs.
+
+## Architecture
+
+- 25 ms causal analysis frame and 10 ms hop;
+- 40 log-mel bands plus energy, zero-crossing rate, and spectral flatness;
+- four causal depthwise temporal blocks with dilations 1, 2, 4, and 8;
+- one 64-unit GRU;
+- speech and auxiliary event heads;
+- 184,680 bytes of FP32 parameters.
+
+## Training inputs and provenance limit
+
+Historical training notes associated with the retained checkpoint report:
+
+- 558 derived clips from nine FLEURS configurations: Arabic, English,
+  Gujarati, Hindi, Kannada, Punjabi, Tamil, Telugu, and Urdu;
+- 288 AMI meeting clips with meeting-family-disjoint train/validation splits;
+- 64 MUSAN noise clips;
+- weak frame targets from the official Silero VAD model.
+
+FLEURS, AMI, and MUSAN attribution is in `NOTICE`; dataset audio is not
+distributed here.
+
+Those counts and corpus names are not embedded as complete provenance in the
+public checkpoint. The exact retained training manifests, their digests, source
+revisions, and teacher-output digest were not preserved, so the checkpoint's
+training run is **not bit-for-bit reproducible** from the public tree. This is a
+provenance limitation, not evidence of broader accuracy. Future release
+candidates must preserve those records before training begins.
+
+## Evaluation
+
+The retained checkpoint was repeatedly inspected on TEN VAD's public
+30-recording set while research candidates were compared. At the retained
+detector policy, the descriptive results are:
+
+- 26,243 frames;
+- ROC-AUC: 0.882;
+- raw F1 at the configured threshold: 0.886;
+- hysteresis-decision F1: 0.889;
+- false-alarm rate: 26.3%;
+- miss rate: 13.0%.
+
+Because the public set influenced research decisions, this is an exploratory
+external-set result—not an untouched test, independent benchmark, or
+production generalization estimate. Language is recorded as `und`, and codec,
+channel, device, and SNR are unknown.
+
+The machine-readable report is
+`benchmarks/flashvad-v0.1/ten-public-evaluation.json`. Its false-alarm rate is
+too high for production.
+
+## Known limitations
+
+- Quiet speech, music, TTS leakage, echo, television, laughter, singing, and
+  overlapping speakers may cause misses or false triggers.
+- Read speech and meetings do not cover real carrier, device, packet-loss, and
+  room conditions.
+- Language presence in training does not prove per-language performance.
+- Linear 8-to-16 kHz conversion prioritizes causal speed, not audio fidelity.
+- VAD cannot determine whether a speaker has semantically completed a turn.
+
+A production candidate needs consented, human-labelled, speaker-disjoint calls
+with predeclared per-slice gates and a test set untouched until final
+evaluation.
+
+## Licences
+
+Repository source code is MIT-licensed. The retained model artifacts are
+separately available under CC BY 4.0; see
+`MODEL_LICENSE.md` and `NOTICE`. Third-party datasets, models, and benchmark
+materials retain their own terms.

SHA256SUMS

@@ -0,0 +1,6 @@
+ca9e35475518466b2a1f2e89b4953cd1e26e3d8c513cdcf265ab319e74e2b288  flashvad-v0.1.pt
+9a88e34bf3118d60e25a16cb622cb394e2f3ab71445b0aa5957df6f1d5f1b6ba  flashvad-stream.onnx
+0b1ad372808f7c67cea5a1ca4b41a817714c9a0b0cf49b3baa56fe8d5f64ad2b  config.json
+b5d000e0406d81fbd87a9e66194a877fa8433a76783faac8275c7969c43051b4  detector-calibration.json
+5052faab08f4a50c9b628b7a57e1747d368b3e7d4cedd3c92da8e3095ba3ea65  flashvad-stream.json
+d4b0097061d3637e40646bec84650d3870b05bbaab2e0f5d4bb8654cc1491f4b  training-history.json

benchmarks/flashvad-v0.1/ten-public-evaluation.json

@@ -0,0 +1,240 @@
+{
+  "schema": "flashvad-exploratory-evaluation-v1",
+  "status": "exploratory_repeated_public_set_evaluation",
+  "disclosure": "This public set was consulted across research candidates. Results are descriptive and must not be presented as an untouched test or production generalization estimate.",
+  "checkpoint": "models/flashvad-v0.1/flashvad-v0.1.pt",
+  "checkpoint_sha256": "ca9e35475518466b2a1f2e89b4953cd1e26e3d8c513cdcf265ab319e74e2b288",
+  "manifest": "TEN public test set (30 files)",
+  "source": {
+    "repository": "https://github.com/TEN-framework/ten-vad",
+    "revision": "22a3bcd4509d0faaa8eef4881e8af5f39c178950",
+    "manifest_digest": null,
+    "reproducibility_limit": "The exact converted manifest and per-file outputs were not retained with this artifact."
+  },
+  "items": 30,
+  "report": {
+    "aggregate": {
+      "frames": 26243,
+      "speech_fraction": 0.7519338726997375,
+      "threshold": 0.49999999999999994,
+      "roc_auc": 0.8819343068676508,
+      "pr_auc": 0.9572105585618752,
+      "accuracy": 0.8310787638608391,
+      "precision": 0.8982300884955752,
+      "recall": 0.8744235544519333,
+      "f1": 0.8861669619700587,
+      "false_alarm_rate": 0.30030721966205837,
+      "miss_rate": 0.12557644554806668,
+      "reference_segments": 126,
+      "predicted_segments": 337,
+      "matched_segments": 121,
+      "segment_recall": 0.9603174603174603,
+      "short_utterance_recall": 1.0,
+      "onset_error_p50_ms": 60.0,
+      "onset_error_p95_ms": 1080.0,
+      "offset_error_p50_ms": 10.0,
+      "offset_error_p95_ms": 330.0,
+      "premature_end_rate_100ms": 0.371900826446281,
+      "false_triggers": 216,
+      "false_triggers_per_noise_hour": 11889.908256880733
+    },
+    "oracle_test_threshold": {
+      "warning": "Diagnostic only. Never select a release threshold on the test set.",
+      "threshold": 0.38999999999999996,
+      "accuracy": 0.8331364554357352,
+      "precision": 0.8757709251101321,
+      "recall": 0.906704505143668,
+      "f1": 0.8909693001020841,
+      "false_alarm_rate": 0.3898617511520737,
+      "miss_rate": 0.09329549485633203
+    },
+    "groups": {
+      "language": {
+        "und": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "domain": {
+        "public-benchmark": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "channel": {
+        "unknown": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "codec": {
+        "unknown": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "device": {
+        "unknown": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "condition": {
+        "ten-public-testset": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      },
+      "snr": {
+        "unknown": {
+          "frames": 26243,
+          "speech_fraction": 0.7519338726997375,
+          "threshold": 0.49999999999999994,
+          "roc_auc": 0.8819343068676508,
+          "pr_auc": 0.9572105585618752,
+          "accuracy": 0.8310787638608391,
+          "precision": 0.8982300884955752,
+          "recall": 0.8744235544519333,
+          "f1": 0.8861669619700587,
+          "false_alarm_rate": 0.30030721966205837,
+          "miss_rate": 0.12557644554806668
+        }
+      }
+    },
+    "confidence_intervals": {
+      "method": "item_bootstrap",
+      "confidence_level": 0.95,
+      "iterations": 1000,
+      "seed": 20260726,
+      "threshold": 0.49999999999999994,
+      "threshold_policy": "fixed_supplied_threshold",
+      "metrics": {
+        "roc_auc": {
+          "lower": 0.8527576527660946,
+          "upper": 0.9096820873574871
+        },
+        "pr_auc": {
+          "lower": 0.9431235727029147,
+          "upper": 0.9695627537195856
+        },
+        "f1": {
+          "lower": 0.8558978786995401,
+          "upper": 0.9091075843752642
+        },
+        "false_alarm_rate": {
+          "lower": 0.23351331173319365,
+          "upper": 0.3659907698080037
+        },
+        "miss_rate": {
+          "lower": 0.08368710277707596,
+          "upper": 0.17908854936116203
+        }
+      },
+      "production_detector_metrics": {
+        "roc_auc": {
+          "lower": 0.770814211710769,
+          "upper": 0.8337388187174674
+        },
+        "pr_auc": {
+          "lower": 0.867475058413576,
+          "upper": 0.9084727586283602
+        },
+        "f1": {
+          "lower": 0.8555711879216696,
+          "upper": 0.9152932721739816
+        },
+        "false_alarm_rate": {
+          "lower": 0.20229136044741164,
+          "upper": 0.327427047690151
+        },
+        "miss_rate": {
+          "lower": 0.07915514707921022,
+          "upper": 0.19416996455779523
+        }
+      }
+    },
+    "production_detector": {
+      "detector": {
+        "start_threshold": 0.8,
+        "stop_threshold": 0.5,
+        "start_frames": 3,
+        "stop_frames": 4,
+        "pre_roll_frames": 3
+      },
+      "frames": 26243,
+      "speech_fraction": 0.7519338726997375,
+      "accuracy": 0.8369469953892467,
+      "precision": 0.9092257176146594,
+      "recall": 0.8700146961941925,
+      "f1": 0.8891881393240968,
+      "false_alarm_rate": 0.26328725038402456,
+      "miss_rate": 0.12998530380580753,
+      "reference_segments": 126,
+      "predicted_segments": 162,
+      "matched_segments": 115,
+      "segment_recall": 0.9126984126984127,
+      "short_utterance_recall": 1.0,
+      "onset_error_p50_ms": 40.0,
+      "onset_error_p95_ms": 1106.999999999999,
+      "offset_error_p50_ms": 40.0,
+      "offset_error_p95_ms": 559.9999999999991,
+      "premature_end_rate_100ms": 0.26956521739130435,
+      "false_triggers": 47,
+      "false_triggers_per_noise_hour": 2587.1559633027523
+    }
+  }
+}

config.json

@@ -0,0 +1,45 @@
+{
+  "seed": 20260726,
+  "feature": {
+    "sample_rate": 16000,
+    "frame_ms": 25.0,
+    "hop_ms": 10.0,
+    "n_fft": 512,
+    "n_mels": 40,
+    "f_min": 50.0,
+    "f_max": 7600.0
+  },
+  "model": {
+    "feature_dim": 43,
+    "hidden_dim": 64,
+    "kernel_size": 3,
+    "dilations": [
+      1,
+      2,
+      4,
+      8
+    ],
+    "recurrent_dim": 64,
+    "dropout": 0.08
+  },
+  "training": {
+    "chunk_seconds": 4.0,
+    "batch_size": 24,
+    "epochs": 20,
+    "learning_rate": 0.001,
+    "weight_decay": 0.0001,
+    "num_workers": 0,
+    "positive_weight": 1.0,
+    "boundary_weight": 0.2,
+    "auxiliary_weight": 0.1,
+    "gradient_clip": 5.0,
+    "detector_max_false_alarm_rate": 0.1
+  },
+  "detector": {
+    "start_threshold": 0.8,
+    "stop_threshold": 0.5,
+    "start_frames": 3,
+    "stop_frames": 4,
+    "pre_roll_frames": 3
+  }
+}

detector-calibration.json

@@ -0,0 +1,19 @@
+{
+  "selection_split": "validation",
+  "max_false_alarm_rate": 0.1,
+  "detector": {
+    "start_threshold": 0.8,
+    "stop_threshold": 0.5,
+    "start_frames": 3,
+    "stop_frames": 4,
+    "pre_roll_frames": 3
+  },
+  "metrics": {
+    "accuracy": 0.9266013824884792,
+    "precision": 0.8959382916346454,
+    "recall": 0.9205954609265325,
+    "f1": 0.908099531193653,
+    "false_alarm_rate": 0.06949513381995134,
+    "miss_rate": 0.07940453907346748
+  }
+}

flashvad-stream.json

@@ -0,0 +1,33 @@
+{
+  "feature": {
+    "sample_rate": 16000,
+    "frame_ms": 25.0,
+    "hop_ms": 10.0,
+    "n_fft": 512,
+    "n_mels": 40,
+    "f_min": 50.0,
+    "f_max": 7600.0
+  },
+  "model": {
+    "feature_dim": 43,
+    "hidden_dim": 64,
+    "kernel_size": 3,
+    "dilations": [
+      1,
+      2,
+      4,
+      8
+    ],
+    "recurrent_dim": 64,
+    "dropout": 0.08
+  },
+  "detector": {
+    "start_threshold": 0.8,
+    "stop_threshold": 0.5,
+    "start_frames": 3,
+    "stop_frames": 4,
+    "pre_roll_frames": 3
+  },
+  "mode": "streaming",
+  "note": "Graph accepts precomputed causal features."
+}

flashvad-stream.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a88e34bf3118d60e25a16cb622cb394e2f3ab71445b0aa5957df6f1d5f1b6ba
+size 208103

flashvad-v0.1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca9e35475518466b2a1f2e89b4953cd1e26e3d8c513cdcf265ab319e74e2b288
+size 195653

training-history.json

@@ -0,0 +1,262 @@
+[
+  {
+    "epoch": 1,
+    "loss": 0.6686453428761713,
+    "elapsed_seconds": 8.71256791599808,
+    "accuracy": 0.8037442396313365,
+    "precision": 0.7224080267558528,
+    "recall": 0.814927468413664,
+    "f1": 0.765883759603095,
+    "false_alarm_rate": 0.2035241788321168,
+    "miss_rate": 0.185072531586336,
+    "threshold": 0.31999999999999995,
+    "fixed_0_5_f1": 0.7043037035788339
+  },
+  {
+    "epoch": 2,
+    "loss": 0.554987029782657,
+    "elapsed_seconds": 7.3522632909989625,
+    "accuracy": 0.846036866359447,
+    "precision": 0.7891653245959904,
+    "recall": 0.8312178287318671,
+    "f1": 0.8096458992108938,
+    "false_alarm_rate": 0.14433166058394162,
+    "miss_rate": 0.16878217126813289,
+    "threshold": 0.4099999999999999,
+    "fixed_0_5_f1": 0.7949665556946948
+  },
+  {
+    "epoch": 3,
+    "loss": 0.5157594269719618,
+    "elapsed_seconds": 7.07947454099849,
+    "accuracy": 0.8560714285714286,
+    "precision": 0.7942396875762753,
+    "recall": 0.8565161441272813,
+    "f1": 0.8242031942587772,
+    "false_alarm_rate": 0.1442176094890511,
+    "miss_rate": 0.14348385587271875,
+    "threshold": 0.3,
+    "fixed_0_5_f1": 0.7878360850000801
+  },
+  {
+    "epoch": 4,
+    "loss": 0.45891835360691463,
+    "elapsed_seconds": 7.055543165999552,
+    "accuracy": 0.8850345622119815,
+    "precision": 0.8367874926975826,
+    "recall": 0.8797379503977538,
+    "f1": 0.8577253739003978,
+    "false_alarm_rate": 0.11152296228710462,
+    "miss_rate": 0.12026204960224614,
+    "threshold": 0.37999999999999995,
+    "fixed_0_5_f1": 0.8469178861172189
+  },
+  {
+    "epoch": 5,
+    "loss": 0.4393227614205459,
+    "elapsed_seconds": 7.123187124998367,
+    "accuracy": 0.8942281105990784,
+    "precision": 0.8420822277539185,
+    "recall": 0.9003275620028077,
+    "f1": 0.8702313813622808,
+    "false_alarm_rate": 0.10973616180048662,
+    "miss_rate": 0.09967243799719233,
+    "threshold": 0.5399999999999999,
+    "fixed_0_5_f1": 0.8690602490367971
+  },
+  {
+    "epoch": 6,
+    "loss": 0.40598839213108195,
+    "elapsed_seconds": 7.728116750000481,
+    "accuracy": 0.9017165898617512,
+    "precision": 0.8575449352096977,
+    "recall": 0.9000058493214788,
+    "f1": 0.8782624826976041,
+    "false_alarm_rate": 0.09717153284671533,
+    "miss_rate": 0.09999415067852128,
+    "threshold": 0.62,
+    "fixed_0_5_f1": 0.8706310447189506
+  },
+  {
+    "epoch": 7,
+    "loss": 0.4059724561099348,
+    "elapsed_seconds": 7.395234250001522,
+    "accuracy": 0.906036866359447,
+    "precision": 0.8673250564334086,
+    "recall": 0.8989822180627047,
+    "f1": 0.8828699448529413,
+    "false_alarm_rate": 0.08937804136253041,
+    "miss_rate": 0.10101778193729527,
+    "threshold": 0.4799999999999999,
+    "fixed_0_5_f1": 0.8817869217210391
+  },
+  {
+    "epoch": 8,
+    "loss": 0.37956942669276533,
+    "elapsed_seconds": 7.249073541999678,
+    "accuracy": 0.9088133640552996,
+    "precision": 0.8779884346500187,
+    "recall": 0.8925479644361254,
+    "f1": 0.8852083363548028,
+    "false_alarm_rate": 0.08061511557177616,
+    "miss_rate": 0.10745203556387459,
+    "threshold": 0.57,
+    "fixed_0_5_f1": 0.8843839358115316
+  },
+  {
+    "epoch": 9,
+    "loss": 0.3707546304012167,
+    "elapsed_seconds": 7.321023499996954,
+    "accuracy": 0.9126152073732718,
+    "precision": 0.8823613945333831,
+    "recall": 0.8978708469817501,
+    "f1": 0.890048561281438,
+    "false_alarm_rate": 0.07780185523114355,
+    "miss_rate": 0.10212915301824989,
+    "threshold": 0.4699999999999999,
+    "fixed_0_5_f1": 0.8892169185349741
+  },
+  {
+    "epoch": 10,
+    "loss": 0.36226719926143514,
+    "elapsed_seconds": 8.081949041999906,
+    "accuracy": 0.908652073732719,
+    "precision": 0.8604880994866445,
+    "recall": 0.9167349087505849,
+    "f1": 0.8877214347413587,
+    "false_alarm_rate": 0.09660127737226278,
+    "miss_rate": 0.08326509124941507,
+    "threshold": 0.5399999999999999,
+    "fixed_0_5_f1": 0.886519357143858
+  },
+  {
+    "epoch": 11,
+    "loss": 0.37090496873033457,
+    "elapsed_seconds": 7.963082500002201,
+    "accuracy": 0.9163940092165899,
+    "precision": 0.8903283772425586,
+    "recall": 0.8984265325222274,
+    "f1": 0.8943591236625663,
+    "false_alarm_rate": 0.07192822384428224,
+    "miss_rate": 0.10157346747777257,
+    "threshold": 0.5499999999999999,
+    "fixed_0_5_f1": 0.8930540242557883
+  },
+  {
+    "epoch": 12,
+    "loss": 0.3591049221055261,
+    "elapsed_seconds": 7.2273035420003,
+    "accuracy": 0.91036866359447,
+    "precision": 0.8642230680125765,
+    "recall": 0.9164424426766495,
+    "f1": 0.8895670688431513,
+    "false_alarm_rate": 0.09357892335766424,
+    "miss_rate": 0.08355755732335049,
+    "threshold": 0.5499999999999999,
+    "fixed_0_5_f1": 0.8886739014609798
+  },
+  {
+    "epoch": 13,
+    "loss": 0.34848099124842674,
+    "elapsed_seconds": 7.2729878749996715,
+    "accuracy": 0.9095276497695852,
+    "precision": 0.8587539840366123,
+    "recall": 0.921970051474029,
+    "f1": 0.8892399260941312,
+    "false_alarm_rate": 0.09855915450121655,
+    "miss_rate": 0.07802994852597099,
+    "threshold": 0.5099999999999999,
+    "fixed_0_5_f1": 0.8888232231556137
+  },
+  {
+    "epoch": 14,
+    "loss": 0.36074137584916477,
+    "elapsed_seconds": 7.317171624999901,
+    "accuracy": 0.9165898617511521,
+    "precision": 0.8765928906773978,
+    "recall": 0.9174075807206364,
+    "f1": 0.8965359551846348,
+    "false_alarm_rate": 0.08394160583941605,
+    "miss_rate": 0.08259241927936359,
+    "threshold": 0.49999999999999994,
+    "fixed_0_5_f1": 0.8965359551846348
+  },
+  {
+    "epoch": 15,
+    "loss": 0.34603760571315373,
+    "elapsed_seconds": 7.179848791001859,
+    "accuracy": 0.9125115207373272,
+    "precision": 0.8778125,
+    "recall": 0.9036909218530651,
+    "f1": 0.8905637537468296,
+    "false_alarm_rate": 0.08175562652068126,
+    "miss_rate": 0.09630907814693496,
+    "threshold": 0.48999999999999994,
+    "fixed_0_5_f1": 0.8904491541339805
+  },
+  {
+    "epoch": 16,
+    "loss": 0.341331767110989,
+    "elapsed_seconds": 7.022008708001522,
+    "accuracy": 0.915184331797235,
+    "precision": 0.8766988655509379,
+    "recall": 0.9131083294337857,
+    "f1": 0.8945332645693658,
+    "false_alarm_rate": 0.08346639294403893,
+    "miss_rate": 0.08689167056621432,
+    "threshold": 0.5599999999999999,
+    "fixed_0_5_f1": 0.8933743612906869
+  },
+  {
+    "epoch": 17,
+    "loss": 0.3445210888467986,
+    "elapsed_seconds": 7.093474582998169,
+    "accuracy": 0.9121428571428571,
+    "precision": 0.8657213656387666,
+    "recall": 0.9196010762751521,
+    "f1": 0.8918481960517358,
+    "false_alarm_rate": 0.09270453163017031,
+    "miss_rate": 0.08039892372484791,
+    "threshold": 0.57,
+    "fixed_0_5_f1": 0.8900419857443751
+  },
+  {
+    "epoch": 18,
+    "loss": 0.33865347658765727,
+    "elapsed_seconds": 7.130151833000127,
+    "accuracy": 0.9156566820276497,
+    "precision": 0.8847178077484752,
+    "recall": 0.903632428638278,
+    "f1": 0.8940750922375751,
+    "false_alarm_rate": 0.07652828467153285,
+    "miss_rate": 0.09636757136172204,
+    "threshold": 0.5599999999999999,
+    "fixed_0_5_f1": 0.8932243456466634
+  },
+  {
+    "epoch": 19,
+    "loss": 0.33929446545140496,
+    "elapsed_seconds": 7.077023041998473,
+    "accuracy": 0.9157834101382488,
+    "precision": 0.885062739930098,
+    "recall": 0.9035446888160973,
+    "f1": 0.8942082259978581,
+    "false_alarm_rate": 0.07626216545012166,
+    "miss_rate": 0.09645531118390267,
+    "threshold": 0.61,
+    "fixed_0_5_f1": 0.8917919367766795
+  },
+  {
+    "epoch": 20,
+    "loss": 0.34355349376283845,
+    "elapsed_seconds": 7.516699709001841,
+    "accuracy": 0.9161520737327189,
+    "precision": 0.88662874217089,
+    "recall": 0.9025503041647169,
+    "f1": 0.8945186817009189,
+    "false_alarm_rate": 0.07500760340632603,
+    "miss_rate": 0.09744969583528311,
+    "threshold": 0.61,
+    "fixed_0_5_f1": 0.892152583184308
+  }
+]