| --- |
| language: |
| - en |
| tags: |
| - security-patch-detection |
| - binary-analysis |
| - graph-neural-network |
| - qwen3 |
| base_model: |
| - Qwen/Qwen3-8B |
| - microsoft/unixcoder-base |
| --- |
| |
| # StriderSPD |
|
|
| Full joint-model checkpoint for identifying security patches from paired pre/post binary control-flow graphs and decompiled pseudo-code. |
|
|
| 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). |
|
|
| ## Downloads |
|
|
| - [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). |
| - [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/`. |
|
|
| 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. |
|
|
| Checkpoint SHA256: |
|
|
| ```text |
| 991045a6bd0c8a30b113b1d2f633ed6d464f48ec06f4325b6adfef27db5e01e7 |
| ``` |
|
|
| Benchmark ZIP SHA256: |
|
|
| ```text |
| dde9462e5e57cc84749562e2064bd682376fb523c1f8e3c879a5de4dcbdd0536 |
| ``` |
|
|
| ## Evaluation |
|
|
| 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. |
|
|
| The archived 1,720 predictions from this exact checkpoint produce the following results: |
|
|
| | Accuracy | Precision | Recall | F1 | |
| | --- | --- | --- | --- | |
| | 0.854070 | 0.822979 | 0.957426 | 0.885126 | |
|
|
| ## Training and use |
|
|
| 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. |
|
|
| 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. |
|
|
