Add VIGIL trigger inference artifacts for ROS integration

PUBLIC_MANIFEST.json

@@ -0,0 +1,186 @@
+{
+  "balanced_manifest_sha256": "549134e307f21470cb942acd44c2c27d2b29fcaa8527b9e7f8e2722e3232b58e",
+  "bundle_dir": "finetune/model_bundles/vigil_latest_optimized_20260626_085405",
+  "code_commit": "132ead0",
+  "dataset_fingerprint": "2b78e211183d47fb",
+  "dataset_zip_sha256": "e2e38518d6725449653138e0ee484c4b5903467e418e8968d4b98ada5fd41701",
+  "fold_sha256": "7c1c65da28f87922f111ee1549b61c053323fc876d2cd26346544de0b37b2a5e",
+  "include_qwen_weights": false,
+  "selected_config": {
+    "fewshot_recipe": "no_adaptation_zero_shot_fallback",
+    "top_k": 1,
+    "variant": "stage2_bce_supcon"
+  },
+  "stage1_checkpoint": {
+    "committed": false,
+    "exists": true,
+    "path": "finetune/model_bundles/vigil_latest_optimized_20260626_085405/stage1/checkpoint_best.pt",
+    "sha256": "5439a089c8aabf3eda0e02e6d44b1a3ab190c9e6b2252be4374231fd3f63a58a"
+  },
+  "stage1_metrics": {
+    "device": "cuda:0",
+    "feature_backend": "official_openwakeword",
+    "input_feature_dim": 96,
+    "official_openwakeword_used": true,
+    "parameters": {
+      "frozen": 0,
+      "total": 56321,
+      "trainable": 56321
+    },
+    "pos_weight": 0.8333333134651184,
+    "status": "ok",
+    "test_metrics": {
+      "defined": false,
+      "n": 0
+    },
+    "test_rows": 0,
+    "theta_1": 0.9973222613334656,
+    "threshold_selection": {
+      "metrics": {
+        "auprc": 0.9999981958576892,
+        "auroc": 0.9999978321193201,
+        "balanced_accuracy": 0.9751344086021505,
+        "f1": 0.9745003445899381,
+        "false_positive_rate": 0.0,
+        "false_reject_rate": 0.04973118279569892,
+        "fn": 37,
+        "fp": 0,
+        "n": 1364,
+        "negative": 620,
+        "positive": 744,
+        "precision": 1.0,
+        "recall": 0.9502688172043011,
+        "specificity": 1.0,
+        "threshold": 0.9973222613334656,
+        "tn": 620,
+        "tp": 707
+      },
+      "reason": "met_recall_target_0.95",
+      "threshold": 0.9973222613334656
+    },
+    "train_rows": 1364,
+    "val_rows": 1364
+  },
+  "stage1_threshold": {
+    "metrics": {
+      "auprc": 0.9999981958576892,
+      "auroc": 0.9999978321193201,
+      "balanced_accuracy": 0.9751344086021505,
+      "f1": 0.9745003445899381,
+      "false_positive_rate": 0.0,
+      "false_reject_rate": 0.04973118279569892,
+      "fn": 37,
+      "fp": 0,
+      "n": 1364,
+      "negative": 620,
+      "positive": 744,
+      "precision": 1.0,
+      "recall": 0.9502688172043011,
+      "specificity": 1.0,
+      "threshold": 0.9973222613334656,
+      "tn": 620,
+      "tp": 707
+    },
+    "reason": "met_recall_target_0.95",
+    "threshold": 0.9973222613334656
+  },
+  "stage2_checkpoint": {
+    "committed": false,
+    "exists": true,
+    "path": "finetune/model_bundles/vigil_latest_optimized_20260626_085405/stage2_bce_supcon/checkpoint_best.pt",
+    "sha256": "c5850274a752db91b529119a65fcc19ff90d60a682c2c78f46d7b5b3d0062efb"
+  },
+  "stage2_metrics": {
+    "device": "cuda:0",
+    "input_feature_dim": 2048,
+    "lambda_supcon": 0.1,
+    "no_pair_batches": 156,
+    "qwen_integrity": {
+      "checksums_unchanged": true,
+      "qwen_parameters_modified": false,
+      "representative_checksums_after": {
+        "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+        "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+        "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+        "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+        "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+        "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+        "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+        "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+      },
+      "representative_checksums_before": {
+        "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+        "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+        "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+        "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+        "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+        "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+        "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+        "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+      },
+      "status": "ok",
+      "total_qwen_parameters": 2038052480,
+      "trainable_qwen_parameters": 0
+    },
+    "status": "ok",
+    "test_metrics": {
+      "defined": false,
+      "n": 0
+    },
+    "test_rows": 0,
+    "theta_2": 0.9877771735191345,
+    "threshold_selection": {
+      "metrics": {
+        "auprc": 1.0,
+        "auroc": 1.0,
+        "balanced_accuracy": 0.950268817204301,
+        "f1": 0.9476661951909477,
+        "false_positive_rate": 0.0,
+        "false_reject_rate": 0.09946236559139784,
+        "fn": 74,
+        "fp": 0,
+        "n": 1364,
+        "negative": 620,
+        "positive": 744,
+        "precision": 1.0,
+        "recall": 0.9005376344086021,
+        "specificity": 1.0,
+        "threshold": 0.9877771735191345,
+        "tn": 620,
+        "tp": 670
+      },
+      "reason": "met_recall_target_0.9",
+      "threshold": 0.9877771735191345
+    },
+    "train_rows": 1364,
+    "val_rows": 1364,
+    "variant": "bce_supcon",
+    "verifier_parameters": 561922,
+    "verifier_trainable_parameters": 561922
+  },
+  "stage2_threshold": {
+    "metrics": {
+      "auprc": 1.0,
+      "auroc": 1.0,
+      "balanced_accuracy": 0.950268817204301,
+      "f1": 0.9476661951909477,
+      "false_positive_rate": 0.0,
+      "false_reject_rate": 0.09946236559139784,
+      "fn": 74,
+      "fp": 0,
+      "n": 1364,
+      "negative": 620,
+      "positive": 744,
+      "precision": 1.0,
+      "recall": 0.9005376344086021,
+      "specificity": 1.0,
+      "threshold": 0.9877771735191345,
+      "tn": 620,
+      "tp": 670
+    },
+    "reason": "met_recall_target_0.9",
+    "threshold": 0.9877771735191345
+  },
+  "status": "trained_deployment_candidate_not_scientific_test",
+  "training_policy": "All balanced windows duplicated as train and val for deployment-only final fit/calibration; not a held-out scientific estimate."
+}

README.md

@@ -0,0 +1,64 @@
+---
+license: other
+private: true
+tags:
+- vigil
+- wake-word
+- inference-artifacts
+- ros
+- qwen-asr
+---
+
+# VIGIL Trigger Inference Artifacts
+
+Private integration bundle for the VIGIL two-stage trigger used by the ROS `medical_agent` integration.
+
+This repository contains only the small VIGIL inference heads and public configuration needed for integration testing. It does not contain raw audio, feature caches, predictions, datasets, Qwen weights, openWakeWord weights, logs, or private participant metadata.
+
+## Contents
+
+```text
+PUBLIC_MANIFEST.json
+stage1/checkpoint_best.pt
+stage1/model_config.json
+stage1/threshold.json
+stage1/metrics.json
+stage2_bce_supcon/checkpoint_best.pt
+stage2_bce_supcon/threshold.json
+stage2_bce_supcon/metrics.json
+stage2_bce_supcon/frozen_qwen_integrity.json
+```
+
+## Selected Runtime Configuration
+
+- Stage 1 feature backend: official openWakeWord speech embedding backbone.
+- Stage 1 trainable head: LayerNorm -> 2-layer GRU -> Linear.
+- Stage 1 threshold: `0.9973222613334656`.
+- Stage 2 variant: `stage2_bce_supcon`.
+- Stage 2 input: frozen Qwen audio encoder features.
+- Stage 2 threshold: `0.9877771735191345`.
+- Qwen model: load separately from `Qwen/Qwen3-ASR-1.7B`.
+- Qwen trainable parameters: zero.
+
+## Intended Integration Shape
+
+The ROS node should run raw audio in parallel with ASR:
+
+```text
+16 kHz microphone audio
+  -> rolling windows
+  -> Stage 1 candidate detector
+  -> Stage 2 verifier
+  -> VIGIL trigger event
+
+ASR transcript
+  -> command/question text after trigger
+```
+
+After VIGIL is confirmed, `medical_agent` can classify the transcript as a command or question and route it to command handling, VQA, or the medical LLM.
+
+## Safety Notes
+
+- Do not treat this repository as a training dataset.
+- Do not publish publicly until the VIGIL team approves sharing trained voice-trigger heads.
+- Do not add raw participant audio, feature caches, predictions, exports, or local logs.

stage1/checkpoint_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5439a089c8aabf3eda0e02e6d44b1a3ab190c9e6b2252be4374231fd3f63a58a
+size 229169

stage1/metrics.json

@@ -0,0 +1,44 @@
+{
+  "device": "cuda:0",
+  "feature_backend": "official_openwakeword",
+  "input_feature_dim": 96,
+  "official_openwakeword_used": true,
+  "parameters": {
+    "frozen": 0,
+    "total": 56321,
+    "trainable": 56321
+  },
+  "pos_weight": 0.8333333134651184,
+  "status": "ok",
+  "test_metrics": {
+    "defined": false,
+    "n": 0
+  },
+  "test_rows": 0,
+  "theta_1": 0.9973222613334656,
+  "threshold_selection": {
+    "metrics": {
+      "auprc": 0.9999981958576892,
+      "auroc": 0.9999978321193201,
+      "balanced_accuracy": 0.9751344086021505,
+      "f1": 0.9745003445899381,
+      "false_positive_rate": 0.0,
+      "false_reject_rate": 0.04973118279569892,
+      "fn": 37,
+      "fp": 0,
+      "n": 1364,
+      "negative": 620,
+      "positive": 744,
+      "precision": 1.0,
+      "recall": 0.9502688172043011,
+      "specificity": 1.0,
+      "threshold": 0.9973222613334656,
+      "tn": 620,
+      "tp": 707
+    },
+    "reason": "met_recall_target_0.95",
+    "threshold": 0.9973222613334656
+  },
+  "train_rows": 1364,
+  "val_rows": 1364
+}

stage1/model_config.json

@@ -0,0 +1,15 @@
+{
+  "allow_acoustic_fallback_when_openwakeword_missing": false,
+  "batch_size": 32,
+  "dropout": 0.1,
+  "early_stopping_patience": 5,
+  "epochs": 20,
+  "feature_backend": "openwakeword",
+  "gradient_clip_norm": 5.0,
+  "gru_hidden_size": 64,
+  "gru_layers": 2,
+  "input_dim": 96,
+  "learning_rate": 0.001,
+  "recall_target": 0.95,
+  "weight_decay": 0.0001
+}

stage1/threshold.json

@@ -0,0 +1,23 @@
+{
+  "metrics": {
+    "auprc": 0.9999981958576892,
+    "auroc": 0.9999978321193201,
+    "balanced_accuracy": 0.9751344086021505,
+    "f1": 0.9745003445899381,
+    "false_positive_rate": 0.0,
+    "false_reject_rate": 0.04973118279569892,
+    "fn": 37,
+    "fp": 0,
+    "n": 1364,
+    "negative": 620,
+    "positive": 744,
+    "precision": 1.0,
+    "recall": 0.9502688172043011,
+    "specificity": 1.0,
+    "threshold": 0.9973222613334656,
+    "tn": 620,
+    "tp": 707
+  },
+  "reason": "met_recall_target_0.95",
+  "threshold": 0.9973222613334656
+}

stage2_bce_supcon/checkpoint_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5850274a752db91b529119a65fcc19ff90d60a682c2c78f46d7b5b3d0062efb
+size 2252485

stage2_bce_supcon/frozen_qwen_integrity.json

@@ -0,0 +1,27 @@
+{
+  "checksums_unchanged": true,
+  "qwen_parameters_modified": false,
+  "representative_checksums_after": {
+    "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+    "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+    "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+    "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+    "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+    "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+    "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+    "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+  },
+  "representative_checksums_before": {
+    "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+    "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+    "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+    "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+    "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+    "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+    "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+    "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+  },
+  "status": "ok",
+  "total_qwen_parameters": 2038052480,
+  "trainable_qwen_parameters": 0
+}

stage2_bce_supcon/metrics.json

@@ -0,0 +1,68 @@
+{
+  "device": "cuda:0",
+  "input_feature_dim": 2048,
+  "lambda_supcon": 0.1,
+  "no_pair_batches": 156,
+  "qwen_integrity": {
+    "checksums_unchanged": true,
+    "qwen_parameters_modified": false,
+    "representative_checksums_after": {
+      "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+      "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+      "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+      "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+      "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+      "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+      "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+      "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+    },
+    "representative_checksums_before": {
+      "thinker.audio_tower.layers.0.self_attn.k_proj.bias": -59.839073181152344,
+      "thinker.audio_tower.layers.0.self_attn.k_proj.weight": -6.748814582824707,
+      "thinker.audio_tower.layers.0.self_attn.out_proj.bias": -0.5344572067260742,
+      "thinker.audio_tower.layers.0.self_attn.out_proj.weight": -2.1157467365264893,
+      "thinker.audio_tower.layers.0.self_attn.q_proj.bias": -57.97587585449219,
+      "thinker.audio_tower.layers.0.self_attn.q_proj.weight": -9.140356063842773,
+      "thinker.audio_tower.layers.0.self_attn.v_proj.bias": -1.118425726890564,
+      "thinker.audio_tower.layers.0.self_attn.v_proj.weight": 16.02427101135254
+    },
+    "status": "ok",
+    "total_qwen_parameters": 2038052480,
+    "trainable_qwen_parameters": 0
+  },
+  "status": "ok",
+  "test_metrics": {
+    "defined": false,
+    "n": 0
+  },
+  "test_rows": 0,
+  "theta_2": 0.9877771735191345,
+  "threshold_selection": {
+    "metrics": {
+      "auprc": 1.0,
+      "auroc": 1.0,
+      "balanced_accuracy": 0.950268817204301,
+      "f1": 0.9476661951909477,
+      "false_positive_rate": 0.0,
+      "false_reject_rate": 0.09946236559139784,
+      "fn": 74,
+      "fp": 0,
+      "n": 1364,
+      "negative": 620,
+      "positive": 744,
+      "precision": 1.0,
+      "recall": 0.9005376344086021,
+      "specificity": 1.0,
+      "threshold": 0.9877771735191345,
+      "tn": 620,
+      "tp": 670
+    },
+    "reason": "met_recall_target_0.9",
+    "threshold": 0.9877771735191345
+  },
+  "train_rows": 1364,
+  "val_rows": 1364,
+  "variant": "bce_supcon",
+  "verifier_parameters": 561922,
+  "verifier_trainable_parameters": 561922
+}

stage2_bce_supcon/threshold.json

@@ -0,0 +1,23 @@
+{
+  "metrics": {
+    "auprc": 1.0,
+    "auroc": 1.0,
+    "balanced_accuracy": 0.950268817204301,
+    "f1": 0.9476661951909477,
+    "false_positive_rate": 0.0,
+    "false_reject_rate": 0.09946236559139784,
+    "fn": 74,
+    "fp": 0,
+    "n": 1364,
+    "negative": 620,
+    "positive": 744,
+    "precision": 1.0,
+    "recall": 0.9005376344086021,
+    "specificity": 1.0,
+    "threshold": 0.9877771735191345,
+    "tn": 620,
+    "tp": 670
+  },
+  "reason": "met_recall_target_0.9",
+  "threshold": 0.9877771735191345
+}