Upload EMBER2024 ember-ml models

.gitattributes

@@ -33,3 +33,11 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_APK.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_ELF.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_PDF.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_PE.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_Win32.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_Win64.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_all.tl filter=lfs diff=lfs merge=lfs -text
+lightgbm/lightgbm_dotnet.tl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,359 @@
+---
+license: apache-2.0
+datasets:
+  - joyce8/EMBER2024
+language:
+  - en
+tags:
+  - malware-detection
+  - cybersecurity
+  - onnxruntime
+  - lightgbm
+  - pytorch
+  - tabnet
+  - binary-classification
+pipeline_tag: text-classification
+library_name: onnxruntime
+---
+
+# EMBER2024 Malware Detection Models
+
+4종 모델 아키텍처(DNN, TabNet, Hybrid GBDT2NN, LightGBM)를 [EMBER2024](https://huggingface.co/datasets/joyce8/EMBER2024) 데이터셋의 8개 파일 타입 subset 전체에 대해 학습·평가하고 배포 가능한 포맷으로 변환한 모델 컬렉션입니다.
+
+> **학습 환경**: NVIDIA DGX Spark (GB10 Grace Blackwell, 128 GB 통합 메모리, CUDA 13)  
+> **코드**: [github.com/evan0416/ember-ml](https://github.com/evan0416/ember-ml)  
+> **데이터셋 논문**: [Joyce et al., KDD 2025 (arXiv:2506.05074)](https://arxiv.org/abs/2506.05074)
+
+---
+
+## 모델 구성
+
+| 디렉토리 | 아키텍처 | 배포 포맷 | 파라미터 |
+|---------|---------|---------|---------|
+| `dnn/` | Feed-Forward DNN (PReLU + Dropout) | ONNX (INT8 Static / FP32) | 13.2 M (PE) / 0.73 M (non-PE) |
+| `tabnet/` | TabNet ([Arik & Pfister, 2021](https://arxiv.org/abs/1908.07442)) | ONNX FP32 | ~3 M |
+| `hybrid/` | GBDT2NN ([DeepGBM, KDD 2019](https://www.microsoft.com/en-us/research/publication/deepgbm-a-deep-learning-framework-distilled-by-gbdt-for-online-prediction-tasks/)) | ONNX (nn_part) + LightGBM booster | ~1 M NN |
+| `lightgbm/` | LightGBM (사전학습, [joyce8/EMBER2024-benchmark-models](https://huggingface.co/joyce8/EMBER2024-benchmark-models)) | Treelite `.tl` | — |
+
+### Subset 목록
+
+| Subset | 대상 파일 타입 | 입력 차원 |
+|--------|-------------|---------|
+| `PE` | PE 바이너리 전체 (Win32 + Win64 + .NET) | 2,568 |
+| `Win32` | Windows 32-bit PE | 2,568 |
+| `Win64` | Windows 64-bit PE | 2,568 |
+| `.NET` | .NET 어셈블리 | 2,568 |
+| `APK` | Android APK | 696 |
+| `ELF` | Linux ELF | 696 |
+| `PDF` | PDF 문서 | 696 |
+| `all` | 전체 파일 타입 혼합 | 2,568 |
+
+---
+
+## 디렉토리 구조
+
+파일명 규칙: `{모델}_{subset}[_suffix].{ext}`  
+`.NET` subset은 파일명에서 `dotnet`으로 변환됩니다.
+
+```
+dnn/
+├── dnn_PE.onnx              # INT8 Static (배포용, PE/Win32/Win64/dotnet/all)
+├── dnn_PE_fp32.onnx         # FP32 ONNX   (참조용, INT8 계열만 추가 포함)
+├── dnn_PE.pt                # PyTorch 체크포인트
+├── dnn_PE_metrics.json      # 평가 결과 (AUC, TPR@1%FPR)
+├── dnn_PE_benchmark.json    # 크기·레이턴시
+├── dnn_APK.onnx             # FP32 (non-PE — INT8 AUC 손실 과다)
+├── dnn_APK.pt
+└── ...
+
+tabnet/
+├── tabnet_PE.onnx           # FP32 ONNX (134 MB — sparsemax 언폴딩)
+├── tabnet_PE.zip            # pytorch-tabnet 네이티브 (7 MB, 경량)
+└── ...
+
+hybrid/
+├── hybrid_PE_nnpart.onnx    # GBDT2NN nn_part ONNX (5.1 MB)
+├── hybrid_PE_lgbm.model     # LightGBM booster (3.6 MB)
+├── hybrid_PE.pt             # PyTorch 체크포인트
+└── ...
+
+lightgbm/
+├── lightgbm_PE.tl           # Treelite 직렬화 (플랫폼 독립, 재컴파일 필요)
+└── ...
+```
+
+---
+
+## 성능 결과 (EMBER2024 test set)
+
+> 평가 기준: ROC-AUC, TPR @ 1% FPR (논문 §4.1), challenge set detection rate @ FPR=1% 임계값  
+> Challenge set: 6,315 evasive malware (양성 only, Win32 3,225 / .NET 829 / Win64 814 / PDF 805 / ELF 386 / APK 256)
+
+### DNN
+
+| Subset | ROC-AUC | TPR@1%FPR | 배포 포맷 | 크기 |
+|--------|---------|----------|---------|------|
+| PE | 0.9969 | 0.9539 | INT8 Static ONNX | 13.3 MB |
+| Win32 | 0.9966 | 0.9468 | INT8 Static ONNX | 13.3 MB |
+| Win64 | 0.9969 | 0.9656 | INT8 Static ONNX | 13.3 MB |
+| .NET | 0.9939 | 0.8976 | INT8 Static ONNX | 13.3 MB |
+| all | 0.9942 | 0.9103 | INT8 Static ONNX | 13.3 MB |
+| APK | 0.9761 | 0.7682 | FP32 ONNX | 3.9 MB |
+| ELF | 0.9840 | 0.8103 | FP32 ONNX | 3.9 MB |
+| PDF | 0.9795 | 0.8902 | FP32 ONNX | 3.9 MB |
+
+> non-PE(APK/ELF/PDF)는 입력 696-dim, 파라미터 수 부족으로 INT8 AUC 손실이 크므로 FP32 유지.
+
+### TabNet
+
+| Subset | ROC-AUC | TPR@1%FPR | 배포 포맷 | 크기 |
+|--------|---------|----------|---------|------|
+| PE | 0.9951 | 0.9212 | FP32 ONNX | 134 MB |
+| Win32 | 0.9949 | 0.9322 | FP32 ONNX | 134 MB |
+| Win64 | 0.9946 | 0.9323 | FP32 ONNX | 134 MB |
+| .NET | 0.9925 | 0.8702 | FP32 ONNX | 134 MB |
+| all | 0.9922 | 0.8912 | FP32 ONNX | 134 MB |
+| APK | 0.9741 | 0.7028 | FP32 ONNX | 13.5 MB |
+| ELF | 0.9793 | 0.5460 | FP32 ONNX | 13.5 MB |
+| PDF | 0.9810 | 0.8597 | FP32 ONNX | 13.5 MB |
+
+> PE 계열 ONNX 크기 134 MB: sparsemax attention loop이 ONNX 그래프로 언폴딩되는 구조적 특성. 크기 우선이면 `tabnet_model.zip`(7 MB) 직접 사용 권장.
+
+### Hybrid (GBDT2NN)
+
+| Subset | ROC-AUC | TPR@1%FPR | 배포 포맷 | 크기 |
+|--------|---------|----------|---------|------|
+| PE | 0.9982 | 0.9752 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+| Win32 | 0.9981 | 0.9747 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+| Win64 | 0.9983 | 0.9812 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+| .NET | 0.9961 | 0.9466 | nn_part ONNX + LightGBM booster | 5.1 + 3.5 MB |
+| all | 0.9970 | 0.9528 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+| APK | 0.9821 | 0.8003 | nn_part ONNX + LightGBM booster | 5.1 + 3.5 MB |
+| ELF | 0.9899 | 0.8827 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+| PDF | 0.9879 | 0.9283 | nn_part ONNX + LightGBM booster | 5.1 + 3.6 MB |
+
+### LightGBM (Treelite 컴파일)
+
+| Subset | ROC-AUC | TPR@1%FPR | 크기 (.tl) | 크기 (원본 .model) |
+|--------|---------|----------|-----------|-----------------|
+| PE | 0.9983 | 0.9707 | 5.3 MB | 3.6 MB |
+| Win32 | 0.9984 | 0.9736 | 5.3 MB | 3.6 MB |
+| Win64 | 0.9989 | 0.9831 | 5.3 MB | 3.6 MB |
+| .NET | 0.9980 | 0.9566 | 5.3 MB | 3.5 MB |
+| all | 0.9968 | 0.9440 | 5.3 MB | 3.6 MB |
+| APK | 0.9861 | 0.8157 | 5.3 MB | 3.5 MB |
+| ELF | 0.9929 | 0.9140 | 5.3 MB | 3.6 MB |
+| PDF | 0.9913 | 0.9275 | 5.3 MB | 3.6 MB |
+
+> 원본 LightGBM 모델: [joyce8/EMBER2024-benchmark-models](https://huggingface.co/joyce8/EMBER2024-benchmark-models). `.tl`은 Treelite 3.9.1로 직렬화된 플랫폼 독립 파일 — 각 플랫폼에서 재컴파일 필요.
+
+### Challenge Set Detection Rate
+
+> Challenge set: 6,315 evasive malware (전부 양성). test set FPR=1% 임계값 적용.
+
+| Subset | DNN | TabNet | Hybrid | LightGBM |
+|--------|-----|--------|--------|----------|
+| `.NET` | 58.6% | 70.0% | 80.6% | 79.6% |
+| `APK`  | 27.3% | 29.3% | 34.4% | 33.6% |
+| `ELF`  | 11.7% |  4.4% | 23.8% | 30.3% |
+| `PDF`  | 41.5% | 40.1% | 56.9% | 57.1% |
+| `PE`   | 38.5% | 36.9% | 58.2% | 58.8% |
+| `Win32`| 36.6% | 45.3% | 58.4% | 69.9% |
+| `Win64`| 46.3% | 44.1% | 59.5% | 59.7% |
+| `all`  | 35.3% | 42.3% | 54.1% | 48.4% |
+
+---
+
+## 추론 성능 (Apple M1, darwin-arm64)
+
+> `warm_batch1` 레이턴시: 배치 크기=1, 캐시 웜업 후 측정. 배포 환경(x86_64 Linux)과 다를 수 있음.
+
+### 레이턴시 (ms, warm batch=1)
+
+| Subset | DNN | TabNet | Hybrid | LightGBM |
+|--------|-----|--------|--------|----------|
+| `.NET` | 0.248 | 5.465 | 0.151 | 0.050 |
+| `APK`  | 0.035 | 0.846 | 0.145 | 0.031 |
+| `ELF`  | 0.039 | 0.505 | 0.160 | 0.036 |
+| `PDF`  | 0.036 | 2.230 | 0.172 | 0.048 |
+| `PE`   | 0.290 | 4.402 | 0.138 | 0.028 |
+| `Win32`| 0.288 | 4.693 | 0.141 | 0.044 |
+| `Win64`| 0.220 | 5.621 | 0.422 | 0.039 |
+| `all`  | 0.254 | 4.788 | 0.147 | 0.068 |
+
+> TabNet 레이턴시 높음: sparsemax attention이 ONNX 그래프로 언폴딩되는 구조적 특성.  
+> Hybrid = nn_part ONNX 추론만 측정 (LightGBM leaf extraction 제외).  
+> LightGBM 레이턴시 = 컴파일 `.dylib` 기준; 업로드 파일은 `.tl` (재컴파일 필요).
+
+### 모델 파일 크기 (배포 포맷)
+
+| Subset | DNN | TabNet `.onnx` | TabNet `.zip` | Hybrid (nn+lgbm) | LightGBM `.tl` |
+|--------|-----|----------------|---------------|-------------------|----------------|
+| PE 계열 | 13.3 MB (INT8) | 140.2 MB | 7.4 MB | 5.3 + 3.8 MB | 5.3 MB |
+| non-PE  |  3.9 MB (FP32) |  13.5 MB | 3.2 MB | 5.3 + 3.7 MB | 5.3 MB |
+
+---
+
+## 사용 방법
+
+### 의존성 설치
+
+```bash
+pip install onnxruntime>=1.20 numpy
+# LightGBM / Hybrid 추론 시
+pip install "treelite==3.9.1" "treelite_runtime==3.9.1" lightgbm>=4.6
+# TabNet 체크포인트 직접 사용 시
+pip install pytorch-tabnet>=4.1
+```
+
+### DNN 추론 (ONNX Runtime)
+
+```python
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+# PE subset — INT8 Static
+model_path = hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="dnn/dnn_PE.onnx",
+)
+sess = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
+
+# X: np.ndarray shape (N, 2568), dtype float32
+X = np.random.randn(1, 2568).astype(np.float32)
+logit = sess.run(["logit"], {"features": X})[0]          # shape (N, 1)
+prob  = 1 / (1 + np.exp(-logit.ravel()))                  # sigmoid → [0, 1]
+print(f"malware probability: {prob[0]:.4f}")
+```
+
+```python
+# APK subset — FP32
+model_path = hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="dnn/dnn_APK.onnx",
+)
+sess = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
+X = np.random.randn(1, 696).astype(np.float32)          # non-PE: dim=696
+prob = 1 / (1 + np.exp(-sess.run(["logit"], {"features": X})[0].ravel()))
+```
+
+### TabNet 추론 (ONNX Runtime)
+
+```python
+import numpy as np
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+model_path = hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="tabnet/tabnet_PE.onnx",
+)
+sess = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
+X = np.random.randn(1, 2568).astype(np.float32)
+# 출력: logit (sigmoid 전)
+logit = sess.run(["logit"], {"features": X})[0]
+prob  = 1 / (1 + np.exp(-logit.ravel()))
+```
+
+### Hybrid 추론 (ONNX + LightGBM)
+
+```python
+import numpy as np
+import lightgbm as lgb
+import onnxruntime as ort
+from huggingface_hub import hf_hub_download
+
+# 1. LightGBM booster로 leaf indices 추출
+booster = lgb.Booster(model_file=hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="hybrid/hybrid_PE_lgbm.model",
+))
+X_raw = np.random.randn(1, 2568).astype(np.float64)
+leaf_indices = booster.predict(X_raw, pred_leaf=True).astype(np.int64)  # (N, n_trees)
+
+# 2. GBDT2NN ONNX로 최종 분류
+nn_sess = ort.InferenceSession(hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="hybrid/hybrid_PE_nnpart.onnx",
+), providers=["CPUExecutionProvider"])
+logit = nn_sess.run(["logit"], {"leaf_indices": leaf_indices})[0]
+prob  = 1 / (1 + np.exp(-logit.ravel()))
+print(f"malware probability: {prob[0]:.4f}")
+```
+
+### LightGBM 추론 (Treelite 컴파일 — 빠른 추론)
+
+```python
+# 1. Treelite .tl → 플랫폼별 공유 라이브러리 컴파일 (최초 1회)
+import treelite, treelite_runtime, sys, numpy as np
+from huggingface_hub import hf_hub_download
+
+tl_path = hf_hub_download(
+    repo_id="cycloevan/ember-model",
+    filename="lightgbm/lightgbm_PE.tl",
+)
+tl_model = treelite.Model.deserialize(tl_path)
+lib_ext   = ".dylib" if sys.platform == "darwin" else ".so"
+lib_path  = tl_path.replace(".tl", lib_ext)
+tl_model.export_lib(
+    toolchain="clang" if sys.platform == "darwin" else "gcc",
+    libpath=lib_path,
+    verbose=False,
+)
+
+# 2. 추론
+predictor = treelite_runtime.Predictor(lib_path, verbose=False)
+X = np.random.randn(1, 2568).astype(np.float32)
+prob = predictor.predict(treelite_runtime.DMatrix(X))
+print(f"malware probability: {prob[0]:.4f}")
+```
+
+> **주의**: `treelite==3.9.1` + `treelite_runtime==3.9.1` 필요. 4.x는 `export_lib()` 미지원.
+
+---
+
+## 학습 및 평가 환경
+
+| 항목 | 내용 |
+|------|------|
+| 데이터셋 | [EMBER2024](https://huggingface.co/datasets/joyce8/EMBER2024) — train 52주(2.6 M), test 12주(606 K), challenge 6,315 |
+| Feature 차원 | PE 2,568 (v3) / non-PE 696 (유효 prefix) |
+| Split 정책 | 시간적 순서 고정 (temporal split), 임의 셔플 없음 |
+| 학습 환경 | DGX Spark (GB10 Grace Blackwell, 128 GB, CUDA 13) |
+| 프레임워크 | PyTorch 2.11.0, pytorch-tabnet 4.1, LightGBM 4.6 |
+| 재현 시드 | 42 |
+| DNN 아키텍처 | Linear(2568→2568→1024→512→1) + PReLU + Dropout(0.5) |
+| Hybrid | LightGBM leaf extraction → Linear(n_trees→512→256→1) + PReLU |
+| 평가 지표 | ROC-AUC, PR-AUC, **TPR @ 1% FPR** (논문 §4.1) |
+
+---
+
+## 알려진 한계
+
+- **TabNet ONNX 크기**: sparsemax attention loop 언폴딩으로 PE 계열 ONNX가 134 MB로 팽창. 원본 `tabnet_model.zip`(7 MB)이 경량.
+- **Treelite `.dylib`**: Mac ARM64 전용 사전 컴파일 파일. 다른 플랫폼은 `.tl`에서 재컴파일 필요.
+- **DNN non-PE INT8**: 696-dim 모델은 양자화 AUC 손실이 크므로 FP32 유지.
+- **Hybrid 추론**: 단일 ONNX 파일이 아님 — LightGBM leaf extraction + nn_part ONNX 2단계.
+- **challenge detection rate**: test set에서 FPR=1% 임계값으로 측정. subset별 분포 차이로 값이 상이할 수 있음.
+
+---
+
+## 인용
+
+```bibtex
+@inproceedings{joyce2025ember2024,
+  title     = {EMBER2024: An Open Dataset for Training Behavioral Malware Detection Models},
+  author    = {Joyce, Ruby and Rudd, Ethan M. and others},
+  booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
+  year      = {2025},
+  url       = {https://arxiv.org/abs/2506.05074}
+}
+```
+
+---
+
+## 라이선스
+
+코드 및 모델 가중치: Apache 2.0  
+LightGBM 원본 모델(`hybrid/hybrid_*_lgbm.model`): [joyce8/EMBER2024-benchmark-models](https://huggingface.co/joyce8/EMBER2024-benchmark-models) 라이선스 준수

dnn/dnn_APK.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7fc471fef4a3ab3b1deaeccba5473b81a62a8b4d0c1f435c4238150765cba64
+size 3902529

dnn/dnn_APK.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82128b485920e8532d9ccc8148b52254e61ab091bacb99978f377aa64e54af3d
+size 3929735

dnn/dnn_APK_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "APK",
+  "quant": "fp32",
+  "model_size_mb": 3.903,
+  "original_size_mb": 3.93,
+  "size_reduction_pct": 0.7,
+  "latency_ms": {
+    "cold_batch1": 0.061,
+    "warm_batch1": 0.035,
+    "warm_batch16": 0.215,
+    "warm_batch256": 2.326,
+    "throughput_batch1": 28571.4,
+    "ms_per_sample_batch1": 0.035,
+    "throughput_batch16": 74418.6,
+    "ms_per_sample_batch16": 0.0134,
+    "throughput_batch256": 110060.2,
+    "ms_per_sample_batch256": 0.0091
+  },
+  "original_latency_ms": {
+    "cold_batch1": 0.196,
+    "warm_batch1": 0.094,
+    "warm_batch16": 0.528,
+    "warm_batch256": 1.149,
+    "throughput_batch1": 10638.3,
+    "ms_per_sample_batch1": 0.094,
+    "throughput_batch16": 30303.0,
+    "ms_per_sample_batch16": 0.033,
+    "throughput_batch256": 222802.4,
+    "ms_per_sample_batch256": 0.0045
+  },
+  "test_metrics": {
+    "roc_auc": 0.9760728567708332,
+    "pr_auc": 0.9782617133148918,
+    "tpr_at_1pct_fpr": 0.7681666666666667,
+    "n_samples": 96000,
+    "n_positive": 48000,
+    "auc_delta_vs_original": -0.0
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_APK_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "APK",
+  "quant": "fp32",
+  "test_metrics": {
+    "roc_auc": 0.9760728567708332,
+    "pr_auc": 0.9782617133148918,
+    "tpr_at_1pct_fpr": 0.7681666666666667,
+    "n_samples": 96000,
+    "n_positive": 48000,
+    "auc_delta_vs_original": -0.0
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.2734375,
+    "threshold": 0.893254017132744
+  },
+  "auc_gate_failed": false,
+  "platform": "macOS-15.7.7-arm64-arm-64bit"
+}

dnn/dnn_ELF.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a3ee26c129cfcef77ab60de38f4d5296fd0cff1d7f99d3ca4bbd3bdf2d5e568
+size 3902529

dnn/dnn_ELF.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f875a91e09738ee9371ccde5693f61367f41a7a2cb623096d3e3854da5c32e9c
+size 3929735

dnn/dnn_ELF_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "ELF",
+  "quant": "fp32",
+  "model_size_mb": 3.903,
+  "original_size_mb": 3.93,
+  "size_reduction_pct": 0.7,
+  "latency_ms": {
+    "cold_batch1": 0.055,
+    "warm_batch1": 0.039,
+    "ms_per_sample_batch1": 0.0387,
+    "throughput_batch1": 25864.6,
+    "warm_batch16": 0.183,
+    "ms_per_sample_batch16": 0.0114,
+    "throughput_batch16": 87489.4,
+    "warm_batch256": 2.135,
+    "ms_per_sample_batch256": 0.0083,
+    "throughput_batch256": 119926.0
+  },
+  "original_latency_ms": {
+    "cold_batch1": 0.135,
+    "warm_batch1": 0.094,
+    "ms_per_sample_batch1": 0.0942,
+    "throughput_batch1": 10612.8,
+    "warm_batch16": 0.522,
+    "ms_per_sample_batch16": 0.0326,
+    "throughput_batch16": 30663.9,
+    "warm_batch256": 1.07,
+    "ms_per_sample_batch256": 0.0042,
+    "throughput_batch256": 239209.3
+  },
+  "test_metrics": {
+    "roc_auc": 0.9839878888888888,
+    "pr_auc": 0.9845206683196877,
+    "tpr_at_1pct_fpr": 0.8103333333333333,
+    "n_samples": 12000,
+    "n_positive": 6000,
+    "auc_delta_vs_original": 6e-06
+  },
+  "peak_rss_delta_mb": 2.2,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_ELF_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "ELF",
+  "quant": "fp32",
+  "test_metrics": {
+    "roc_auc": 0.9839878888888888,
+    "pr_auc": 0.9845206683196877,
+    "tpr_at_1pct_fpr": 0.8103333333333333,
+    "n_samples": 12000,
+    "n_positive": 6000,
+    "auc_delta_vs_original": 6e-06
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.11658031088082901,
+    "threshold": 0.9921941479742924
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

dnn/dnn_PDF.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9817bd278a5614376a7c1e1267e4e098cab64a9684936fc847df8b3bae0beac
+size 3902529

dnn/dnn_PDF.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d742551ed4f103eabda97a63ce9fba3be032039b97dd5f8fb3c66a328da7bfa9
+size 3929735

dnn/dnn_PDF_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "PDF",
+  "quant": "fp32",
+  "model_size_mb": 3.903,
+  "original_size_mb": 3.93,
+  "size_reduction_pct": 0.7,
+  "latency_ms": {
+    "cold_batch1": 0.05,
+    "warm_batch1": 0.036,
+    "warm_batch16": 0.201,
+    "warm_batch256": 2.391,
+    "throughput_batch1": 27777.8,
+    "ms_per_sample_batch1": 0.036,
+    "throughput_batch16": 79602.0,
+    "ms_per_sample_batch16": 0.0126,
+    "throughput_batch256": 107068.2,
+    "ms_per_sample_batch256": 0.0093
+  },
+  "original_latency_ms": {
+    "cold_batch1": 0.177,
+    "warm_batch1": 0.102,
+    "warm_batch16": 0.502,
+    "warm_batch256": 1.122,
+    "throughput_batch1": 9803.9,
+    "ms_per_sample_batch1": 0.102,
+    "throughput_batch16": 31872.5,
+    "ms_per_sample_batch16": 0.0314,
+    "throughput_batch256": 228164.0,
+    "ms_per_sample_batch256": 0.0044
+  },
+  "test_metrics": {
+    "roc_auc": 0.9795445,
+    "pr_auc": 0.9848222884024884,
+    "tpr_at_1pct_fpr": 0.8901666666666667,
+    "n_samples": 24000,
+    "n_positive": 12000,
+    "auc_delta_vs_original": 0.0
+  },
+  "peak_rss_delta_mb": 2.1,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_PDF_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "PDF",
+  "quant": "fp32",
+  "test_metrics": {
+    "roc_auc": 0.9795445,
+    "pr_auc": 0.9848222884024884,
+    "tpr_at_1pct_fpr": 0.8901666666666667,
+    "n_samples": 24000,
+    "n_positive": 12000,
+    "auc_delta_vs_original": 0.0
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.41490683229813663,
+    "threshold": 0.7553592072806331
+  },
+  "auc_gate_failed": false,
+  "platform": "macOS-15.7.7-arm64-arm-64bit"
+}

dnn/dnn_PE.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9fbe6a00e369366c42b64aeec8e61bf70d96060fd78575a01354e0027459be7
+size 13303487

dnn/dnn_PE.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:382d174985ddd146df8a9e0d1bae262339734aed7841b35a5000b6b7c4fcdf43
+size 52938695

dnn/dnn_PE_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "PE",
+  "quant": "int8-static",
+  "model_size_mb": 13.303,
+  "original_size_mb": 52.939,
+  "size_reduction_pct": 74.9,
+  "latency_ms": {
+    "cold_batch1": 0.778,
+    "warm_batch1": 0.29,
+    "warm_batch16": 1.036,
+    "warm_batch256": 12.935,
+    "throughput_batch1": 3448.3,
+    "ms_per_sample_batch1": 0.29,
+    "throughput_batch16": 15444.0,
+    "ms_per_sample_batch16": 0.0648,
+    "throughput_batch256": 19791.3,
+    "ms_per_sample_batch256": 0.0505
+  },
+  "original_latency_ms": {
+    "cold_batch1": 2.916,
+    "warm_batch1": 2.466,
+    "warm_batch16": 4.939,
+    "warm_batch256": 11.88,
+    "throughput_batch1": 405.5,
+    "ms_per_sample_batch1": 2.466,
+    "throughput_batch16": 3239.5,
+    "ms_per_sample_batch16": 0.3087,
+    "throughput_batch256": 21548.8,
+    "ms_per_sample_batch256": 0.0464
+  },
+  "test_metrics": {
+    "roc_auc": 0.9968856578492659,
+    "pr_auc": 0.9969432521647822,
+    "tpr_at_1pct_fpr": 0.9471575053488233,
+    "n_samples": 100000,
+    "n_positive": 50011,
+    "auc_delta_vs_original": -6.2e-05
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_PE_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df67ff72e0ee7cfc6381dbec0f8f5945c3db77dadb46af930b3f6c4670a5e150
+size 52851589

dnn/dnn_PE_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "PE",
+  "quant": "int8-static",
+  "test_metrics": {
+    "roc_auc": 0.9968856578492659,
+    "pr_auc": 0.9969432521647822,
+    "tpr_at_1pct_fpr": 0.9471575053488233,
+    "n_samples": 100000,
+    "n_positive": 50011,
+    "auc_delta_vs_original": -6.2e-05
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.3849962862094578,
+    "threshold": 0.9285938677767953
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

dnn/dnn_Win32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:500f4ed1d1317a20b95cc1147f68eecd10abcf43527ad8234d1cba6aa4fa7a28
+size 13303486

dnn/dnn_Win32.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f0f67ab9635cdc54f253a3bc591181c0c40092ddd7613ae48cdad9cf5b633be
+size 52938695

dnn/dnn_Win32_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "Win32",
+  "quant": "int8-static",
+  "model_size_mb": 13.303,
+  "original_size_mb": 52.939,
+  "size_reduction_pct": 74.9,
+  "latency_ms": {
+    "cold_batch1": 0.487,
+    "warm_batch1": 0.288,
+    "warm_batch16": 0.99,
+    "warm_batch256": 8.912,
+    "throughput_batch1": 3472.2,
+    "ms_per_sample_batch1": 0.288,
+    "throughput_batch16": 16161.6,
+    "ms_per_sample_batch16": 0.0619,
+    "throughput_batch256": 28725.3,
+    "ms_per_sample_batch256": 0.0348
+  },
+  "original_latency_ms": {
+    "cold_batch1": 2.393,
+    "warm_batch1": 2.343,
+    "warm_batch16": 4.234,
+    "warm_batch256": 11.15,
+    "throughput_batch1": 426.8,
+    "ms_per_sample_batch1": 2.343,
+    "throughput_batch16": 3778.9,
+    "ms_per_sample_batch16": 0.2646,
+    "throughput_batch256": 22959.6,
+    "ms_per_sample_batch256": 0.0436
+  },
+  "test_metrics": {
+    "roc_auc": 0.9965419976367095,
+    "pr_auc": 0.9966226288532287,
+    "tpr_at_1pct_fpr": 0.9478530944531651,
+    "n_samples": 100000,
+    "n_positive": 50204,
+    "auc_delta_vs_original": -8e-05
+  },
+  "peak_rss_delta_mb": 5.3,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_Win32_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:359d7e4e4cc4222bc8756d179415e93affb84f2602b0587e4f063dbb8047945e
+size 52851589

dnn/dnn_Win32_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "Win32",
+  "quant": "int8-static",
+  "test_metrics": {
+    "roc_auc": 0.9965419976367095,
+    "pr_auc": 0.9966226288532287,
+    "tpr_at_1pct_fpr": 0.9478530944531651,
+    "n_samples": 100000,
+    "n_positive": 50204,
+    "auc_delta_vs_original": -8e-05
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.3655813953488372,
+    "threshold": 0.7420705254159128
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

dnn/dnn_Win64.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24bac8430ecb1b15cf0e28a097eff6fe525025d115b9f1bca9fcf75a31882167
+size 13303486

dnn/dnn_Win64.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39d22c9b0123afb6c0a6c11d17a147905c35b35214bbf318a54acdfff1d892a5
+size 52938695

dnn/dnn_Win64_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "Win64",
+  "quant": "int8-static",
+  "model_size_mb": 13.303,
+  "original_size_mb": 52.939,
+  "size_reduction_pct": 74.9,
+  "latency_ms": {
+    "cold_batch1": 0.426,
+    "warm_batch1": 0.22,
+    "warm_batch16": 0.941,
+    "warm_batch256": 8.919,
+    "throughput_batch1": 4545.5,
+    "ms_per_sample_batch1": 0.22,
+    "throughput_batch16": 17003.2,
+    "ms_per_sample_batch16": 0.0588,
+    "throughput_batch256": 28702.8,
+    "ms_per_sample_batch256": 0.0348
+  },
+  "original_latency_ms": {
+    "cold_batch1": 2.789,
+    "warm_batch1": 2.389,
+    "warm_batch16": 4.246,
+    "warm_batch256": 11.135,
+    "throughput_batch1": 418.6,
+    "ms_per_sample_batch1": 2.389,
+    "throughput_batch16": 3768.3,
+    "ms_per_sample_batch16": 0.2654,
+    "throughput_batch256": 22990.6,
+    "ms_per_sample_batch256": 0.0435
+  },
+  "test_metrics": {
+    "roc_auc": 0.9968797740122215,
+    "pr_auc": 0.9970818991078805,
+    "tpr_at_1pct_fpr": 0.9616965868794326,
+    "n_samples": 100000,
+    "n_positive": 49920,
+    "auc_delta_vs_original": -3.3e-05
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_Win64_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc072a860523d3e1557fcb98837758d5f064d5edb57cec2e98931b5d51928bb6
+size 52851589

dnn/dnn_Win64_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "Win64",
+  "quant": "int8-static",
+  "test_metrics": {
+    "roc_auc": 0.9968797740122215,
+    "pr_auc": 0.9970818991078805,
+    "tpr_at_1pct_fpr": 0.9616965868794326,
+    "n_samples": 100000,
+    "n_positive": 49920,
+    "auc_delta_vs_original": -3.3e-05
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.4631449631449631,
+    "threshold": 0.5563582119497066
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

dnn/dnn_all.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55ffa1127b72aa5f593b4963f769697d6a696beb60995ac771d1a6ebbb34a1db
+size 13303487

dnn/dnn_all.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5563545cd2037339668863a623d0b6f84e90cec5053d85abbc9a4e3faa8ad3e
+size 52938695

dnn/dnn_all_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": "all",
+  "quant": "int8-static",
+  "model_size_mb": 13.303,
+  "original_size_mb": 52.939,
+  "size_reduction_pct": 74.9,
+  "latency_ms": {
+    "cold_batch1": 0.429,
+    "warm_batch1": 0.254,
+    "warm_batch16": 0.93,
+    "warm_batch256": 8.955,
+    "throughput_batch1": 3937.0,
+    "ms_per_sample_batch1": 0.254,
+    "throughput_batch16": 17204.3,
+    "ms_per_sample_batch16": 0.0581,
+    "throughput_batch256": 28587.4,
+    "ms_per_sample_batch256": 0.035
+  },
+  "original_latency_ms": {
+    "cold_batch1": 2.658,
+    "warm_batch1": 2.394,
+    "warm_batch16": 4.271,
+    "warm_batch256": 11.371,
+    "throughput_batch1": 417.7,
+    "ms_per_sample_batch1": 2.394,
+    "throughput_batch16": 3746.2,
+    "ms_per_sample_batch16": 0.2669,
+    "throughput_batch256": 22513.4,
+    "ms_per_sample_batch256": 0.0444
+  },
+  "test_metrics": {
+    "roc_auc": 0.9938331270533947,
+    "pr_auc": 0.9937970701852039,
+    "tpr_at_1pct_fpr": 0.8870437241462001,
+    "n_samples": 100000,
+    "n_positive": 50041,
+    "auc_delta_vs_original": -0.000375
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_all_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c97fe488d5dedac6e88793a271b47018301dd7c7fb59c1cdbd23dad489e86f5b
+size 52851589

dnn/dnn_all_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": "all",
+  "quant": "int8-static",
+  "test_metrics": {
+    "roc_auc": 0.9938331270533947,
+    "pr_auc": 0.9937970701852039,
+    "tpr_at_1pct_fpr": 0.8870437241462001,
+    "n_samples": 100000,
+    "n_positive": 50041,
+    "auc_delta_vs_original": -0.000375
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.3534441805225653,
+    "threshold": 0.9838834411565777
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

dnn/dnn_dotnet.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dff033cb3a75fc3af64c11c6557207974480d7286fea7d50225d5f41c0f871cb
+size 13303487

dnn/dnn_dotnet.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05c84608a67b7a3de08102f0d8b75c2fbd9ae3b5a97224d854717f258dab8abf
+size 52938695

dnn/dnn_dotnet_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "dnn",
+  "subset": ".NET",
+  "quant": "int8-static",
+  "model_size_mb": 13.303,
+  "original_size_mb": 52.939,
+  "size_reduction_pct": 74.9,
+  "latency_ms": {
+    "cold_batch1": 0.429,
+    "warm_batch1": 0.248,
+    "warm_batch16": 0.909,
+    "warm_batch256": 8.858,
+    "throughput_batch1": 4032.3,
+    "ms_per_sample_batch1": 0.248,
+    "throughput_batch16": 17601.8,
+    "ms_per_sample_batch16": 0.0568,
+    "throughput_batch256": 28900.4,
+    "ms_per_sample_batch256": 0.0346
+  },
+  "original_latency_ms": {
+    "cold_batch1": 2.44,
+    "warm_batch1": 2.368,
+    "warm_batch16": 4.208,
+    "warm_batch256": 11.103,
+    "throughput_batch1": 422.3,
+    "ms_per_sample_batch1": 2.368,
+    "throughput_batch16": 3802.3,
+    "ms_per_sample_batch16": 0.263,
+    "throughput_batch256": 23056.8,
+    "ms_per_sample_batch256": 0.0434
+  },
+  "test_metrics": {
+    "roc_auc": 0.9919714812355357,
+    "pr_auc": 0.9900265259892996,
+    "tpr_at_1pct_fpr": 0.8444236421489054,
+    "n_samples": 100000,
+    "n_positive": 49946,
+    "auc_delta_vs_original": -0.001909
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

dnn/dnn_dotnet_fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e182e7c89766c2625d318c55e854a428b7a1bccdf5e71b52d805eb9fc59aa056
+size 52851589

dnn/dnn_dotnet_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "dnn",
+  "subset": ".NET",
+  "quant": "int8-static",
+  "test_metrics": {
+    "roc_auc": 0.9919714812355357,
+    "pr_auc": 0.9900265259892996,
+    "tpr_at_1pct_fpr": 0.8444236421489054,
+    "n_samples": 100000,
+    "n_positive": 49946,
+    "auc_delta_vs_original": -0.001909
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.586248492159228,
+    "threshold": 0.9723500243107374
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

hybrid/hybrid_APK.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b65ff8cb5da0cf77b0b542785967c4bf63d580a1ff8be1513b3517cf834b6653
+size 5331649

hybrid/hybrid_APK_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "hybrid",
+  "subset": "APK",
+  "quant": "fp32",
+  "model_size_mb": 5.326,
+  "original_size_mb": 5.332,
+  "size_reduction_pct": 0.1,
+  "latency_ms": {
+    "cold_batch1": 0.455,
+    "warm_batch1": 0.145,
+    "warm_batch16": 0.805,
+    "warm_batch256": 11.495,
+    "throughput_batch1": 6896.6,
+    "ms_per_sample_batch1": 0.145,
+    "throughput_batch16": 19875.8,
+    "ms_per_sample_batch16": 0.0503,
+    "throughput_batch256": 22270.6,
+    "ms_per_sample_batch256": 0.0449
+  },
+  "original_latency_ms": {
+    "cold_batch1": 0.466,
+    "warm_batch1": 0.211,
+    "warm_batch16": 0.908,
+    "warm_batch256": 6.452,
+    "throughput_batch1": 4739.3,
+    "ms_per_sample_batch1": 0.211,
+    "throughput_batch16": 17621.1,
+    "ms_per_sample_batch16": 0.0568,
+    "throughput_batch256": 39677.6,
+    "ms_per_sample_batch256": 0.0252
+  },
+  "test_metrics": {
+    "roc_auc": 0.9827782760416667,
+    "pr_auc": 0.9845011360409881,
+    "tpr_at_1pct_fpr": 0.8002916666666666,
+    "n_samples": 96000,
+    "n_positive": 48000,
+    "auc_delta_vs_original": 0.000631
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

hybrid/hybrid_APK_lgbm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a13274463201e7d3ddec74b098350051266fcf0812aff302617397f1765cbc90
+size 3679987

hybrid/hybrid_APK_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "hybrid",
+  "subset": "APK",
+  "quant": "fp32",
+  "test_metrics": {
+    "roc_auc": 0.9827782760416667,
+    "pr_auc": 0.9845011360409881,
+    "tpr_at_1pct_fpr": 0.8002916666666666,
+    "n_samples": 96000,
+    "n_positive": 48000,
+    "auc_delta_vs_original": 0.000631
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.34375,
+    "threshold": 0.9999923489401177
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

hybrid/hybrid_APK_nnpart.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7052c06fe40e9b2ab700f004f5d5e95f80e78d644073b13caa64e393ec2ae8e9
+size 5325521

hybrid/hybrid_ELF.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a1e2daf9a192da640e68210bd3a6ecf9031f938e32508cd23365481b1cdc560
+size 5331649

hybrid/hybrid_ELF_benchmark.json

@@ -0,0 +1,43 @@
+{
+  "model": "hybrid",
+  "subset": "ELF",
+  "quant": "fp32",
+  "model_size_mb": 5.326,
+  "original_size_mb": 5.332,
+  "size_reduction_pct": 0.1,
+  "latency_ms": {
+    "cold_batch1": 0.483,
+    "warm_batch1": 0.16,
+    "warm_batch16": 1.058,
+    "warm_batch256": 12.195,
+    "throughput_batch1": 6250.0,
+    "ms_per_sample_batch1": 0.16,
+    "throughput_batch16": 15122.9,
+    "ms_per_sample_batch16": 0.0661,
+    "throughput_batch256": 20992.2,
+    "ms_per_sample_batch256": 0.0476
+  },
+  "original_latency_ms": {
+    "cold_batch1": 0.487,
+    "warm_batch1": 0.211,
+    "warm_batch16": 0.912,
+    "warm_batch256": 6.459,
+    "throughput_batch1": 4739.3,
+    "ms_per_sample_batch1": 0.211,
+    "throughput_batch16": 17543.9,
+    "ms_per_sample_batch16": 0.057,
+    "throughput_batch256": 39634.6,
+    "ms_per_sample_batch256": 0.0252
+  },
+  "test_metrics": {
+    "roc_auc": 0.9899468888888888,
+    "pr_auc": 0.9863592706579489,
+    "tpr_at_1pct_fpr": 0.8826666666666667,
+    "n_samples": 12000,
+    "n_positive": 6000,
+    "auc_delta_vs_original": 0.0
+  },
+  "peak_rss_delta_mb": 0.0,
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "platform_short": "darwin-arm64"
+}

hybrid/hybrid_ELF_lgbm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e4072938955ac6020f80bca6cff7c12e97865102d578b3099e2011048eb9c43
+size 3764582

hybrid/hybrid_ELF_metrics.json

@@ -0,0 +1,19 @@
+{
+  "model": "hybrid",
+  "subset": "ELF",
+  "quant": "fp32",
+  "test_metrics": {
+    "roc_auc": 0.9899468888888888,
+    "pr_auc": 0.9863592706579489,
+    "tpr_at_1pct_fpr": 0.8826666666666667,
+    "n_samples": 12000,
+    "n_positive": 6000,
+    "auc_delta_vs_original": 0.0
+  },
+  "challenge_metrics": {
+    "detection_rate": 0.23834196891191708,
+    "threshold": 0.7018789316233345
+  },
+  "platform": "macOS-15.7.7-arm64-arm-64bit",
+  "auc_gate_failed": false
+}

hybrid/hybrid_ELF_nnpart.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:182798a23e2a8f7c478d561abd8103c32dcb9e8feb6decf0728261ab9d0893c6
+size 5325521

hybrid/hybrid_PDF.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5785b87d151d9f4fefd2184df92444655697fe77fc9db74fa52e787dd8139ff7
+size 5331649