Add files using upload-large-folder tool
Browse files- README.md +57 -0
- SHA256SUMS +2 -0
- benchmark_1720.zip +3 -0
- model.json +16 -0
- model_epoch5_best_acc_qwen3_8b_824.pth +3 -0
- paper_reference_metrics.json +22 -0
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
tags:
|
| 5 |
+
- security-patch-detection
|
| 6 |
+
- binary-analysis
|
| 7 |
+
- graph-neural-network
|
| 8 |
+
- qwen3
|
| 9 |
+
base_model:
|
| 10 |
+
- Qwen/Qwen3-8B
|
| 11 |
+
- microsoft/unixcoder-base
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# StriderSPD
|
| 15 |
+
|
| 16 |
+
Full joint-model checkpoint for identifying security patches from paired pre/post binary control-flow graphs and decompiled pseudo-code.
|
| 17 |
+
|
| 18 |
+
The checkpoint contains 718 tensors, including the Qwen3-8B backbone, UniXcoder node encoder, GNN, cross-attention and binary classifier. Load it with the custom StriderSPD code in [7thParkk/StriderSPD](https://github.com/7thParkk/StriderSPD).
|
| 19 |
+
|
| 20 |
+
## Downloads
|
| 21 |
+
|
| 22 |
+
- [Paper checkpoint](https://huggingface.co/hanjin996/StriderSPD/resolve/main/model_epoch5_best_acc_qwen3_8b_824.pth): 25,682,301,190 bytes (25.68 GB).
|
| 23 |
+
- [1,720-sample benchmark](https://huggingface.co/hanjin996/StriderSPD/resolve/main/benchmark_1720.zip): extract into `benchmark_1720/`. The ZIP contains `benchmark_new_infos.json`, `CFG/` and `pseudo/`.
|
| 24 |
+
|
| 25 |
+
The benchmark is also included directly in the GitHub repository. Each modality contains 3,440 JSON files, representing 1,720 paired pre/post samples. Labels: 1,010 security and 710 non-security.
|
| 26 |
+
|
| 27 |
+
Checkpoint SHA256:
|
| 28 |
+
|
| 29 |
+
```text
|
| 30 |
+
991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7
|
| 31 |
+
```
|
| 32 |
+
|
| 33 |
+
Benchmark ZIP SHA256:
|
| 34 |
+
|
| 35 |
+
```text
|
| 36 |
+
dde9462e5e57cc84749562e2064bd682376fb523c1f8e3c879a5de4dcbdd0536
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
## Evaluation
|
| 40 |
+
|
| 41 |
+
See the GitHub README for environment installation and commands. The supported environment is Linux, Python 3.11, CUDA and BF16-capable GPUs. The original execution used four visible GPUs. Evaluation initializes the custom architecture using Qwen3-8B and UniXcoder, then strictly loads all checkpoint tensors.
|
| 42 |
+
|
| 43 |
+
The archived 1,720 predictions from this exact checkpoint produce the following results:
|
| 44 |
+
|
| 45 |
+
| Accuracy | Precision | Recall | F1 |
|
| 46 |
+
| --- | --- | --- | --- |
|
| 47 |
+
| 0.854070 | 0.822979 | 0.957426 | 0.885126 |
|
| 48 |
+
|
| 49 |
+
Confusion matrix: TP 967, TN 502, FP 208, FN 43. Positive means a security patch; the decision threshold is sigmoid(logit) >= 0.5. The reference prediction records and metrics are available in the GitHub `release_metadata/` directory.
|
| 50 |
+
|
| 51 |
+
The local checkpoint checksum and archived predictions have been verified. GPU inference with the portable release has not been rerun. The archived evaluation used a local LoRA-merged Qwen3 initializer; its tokenizer/config bytes have not been compared with the public model assets. All saved model weights are included in the joint checkpoint.
|
| 52 |
+
|
| 53 |
+
## Training and use
|
| 54 |
+
|
| 55 |
+
The GitHub code exposes the joint-training stage. Paper training initializes from a LoRA-merged Qwen3-8B model and requires the prepared train/dev data. The released 1,720-sample benchmark is for evaluation.
|
| 56 |
+
|
| 57 |
+
Model outputs provide a signal for prioritizing patch review. Evaluate performance on the intended projects and binary-analysis pipeline before use. The Qwen3-8B and UniXcoder upstream model licenses and conditions also apply to their components.
|
SHA256SUMS
ADDED
|
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
|
|
|
| 1 |
+
991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7 model_epoch5_best_acc_qwen3_8b_824.pth
|
| 2 |
+
dde9462e5e57cc84749562e2064bd682376fb523c1f8e3c879a5de4dcbdd0536 benchmark_1720.zip
|
benchmark_1720.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dde9462e5e57cc84749562e2064bd682376fb523c1f8e3c879a5de4dcbdd0536
|
| 3 |
+
size 23732628
|
model.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "StriderSPD",
|
| 3 |
+
"repository": "hanjin996/StriderSPD",
|
| 4 |
+
"filename": "model_epoch5_best_acc_qwen3_8b_824.pth",
|
| 5 |
+
"size_bytes": 25682301190,
|
| 6 |
+
"sha256": "991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7",
|
| 7 |
+
"download_url": "https://huggingface.co/hanjin996/StriderSPD/resolve/main/model_epoch5_best_acc_qwen3_8b_824.pth",
|
| 8 |
+
"format": "PyTorch full joint-model state_dict",
|
| 9 |
+
"state_dict_entries": 718,
|
| 10 |
+
"llm": "Qwen3-8B",
|
| 11 |
+
"node_encoder": "microsoft/unixcoder-base",
|
| 12 |
+
"attention_layer_frequency": 1,
|
| 13 |
+
"reference_metrics": "release_metadata/paper_reference_metrics.json",
|
| 14 |
+
"source_code_commit": "8262094",
|
| 15 |
+
"verification": "Local checkpoint SHA256 matches the server source digest; archived prediction records reproduce the reported benchmark metrics. GPU evaluation of this portable release has not been rerun."
|
| 16 |
+
}
|
model_epoch5_best_acc_qwen3_8b_824.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7
|
| 3 |
+
size 25682301190
|
paper_reference_metrics.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"num_samples": 1720,
|
| 3 |
+
"accuracy": 0.8540697674418605,
|
| 4 |
+
"precision": 0.8229787234042554,
|
| 5 |
+
"recall": 0.9574257425742574,
|
| 6 |
+
"f1": 0.8851258581235698,
|
| 7 |
+
"fpr": 0.29295770521722464,
|
| 8 |
+
"tp": 967,
|
| 9 |
+
"tn": 502,
|
| 10 |
+
"fp": 208,
|
| 11 |
+
"fn": 43,
|
| 12 |
+
"pred_pos_rate": 0.6831395348837209,
|
| 13 |
+
"label_pos_rate": 0.5872093023255814,
|
| 14 |
+
"mean_logit": -0.1554695442665455,
|
| 15 |
+
"prediction_file": "release_metadata/reference_predictions.jsonl",
|
| 16 |
+
"benchmark_manifest": "benchmark_1720/benchmark_new_infos.json",
|
| 17 |
+
"checkpoint": "model_epoch5_best_acc_qwen3_8b_824.pth",
|
| 18 |
+
"code_root": "Original snapshot dated 2026-07-19",
|
| 19 |
+
"reference_export_date": "2026-08-12",
|
| 20 |
+
"decision_rule": "sigmoid(logit) >= 0.5",
|
| 21 |
+
"fpr_denominator_adjustment": 0.0001
|
| 22 |
+
}
|