---
license: mit
datasets:
- mahdin70/cwe_enriched_balanced_bigvul_primevul
metrics:
- accuracy
- precision
- recall
- f1
base_model:
- microsoft/codebert-base
library_name: transformers
---


# CodeBERT-VulnCWE - Fine-Tuned CodeBERT for Vulnerability and CWE Classification

## Model Overview
This model is a fine-tuned version of **microsoft/codebert-base** on a curated and enriched dataset for vulnerability detection and CWE classification. It is capable of predicting whether a given code snippet is vulnerable and, if vulnerable, identifying the specific CWE ID associated with it.

## Dataset
The model was fine-tuned using the dataset [mahdin70/cwe_enriched_balanced_bigvul_primevul](https://huggingface.co/datasets/mahdin70/cwe_enriched_balanced_bigvul_primevul). The dataset contains both vulnerable and non-vulnerable code samples and is enriched with CWE metadata.

### CWE IDs Covered:
1. **CWE-119**: Improper Restriction of Operations within the Bounds of a Memory Buffer  
2. **CWE-20**: Improper Input Validation  
3. **CWE-125**: Out-of-bounds Read  
4. **CWE-399**: Resource Management Errors  
5. **CWE-200**: Information Exposure  
6. **CWE-787**: Out-of-bounds Write  
7. **CWE-264**: Permissions, Privileges, and Access Controls  
8. **CWE-416**: Use After Free  
9. **CWE-476**: NULL Pointer Dereference  
10. **CWE-190**: Integer Overflow or Wraparound  
11. **CWE-189**: Numeric Errors  
12. **CWE-362**: Concurrent Execution using Shared Resource with Improper Synchronization  

---

## Model Training
The model was trained for **3 epochs** with the following configuration:
- **Learning Rate**: 2e-5  
- **Weight Decay**: 0.01  
- **Batch Size**: 8  
- **Optimizer**: AdamW  
- **Scheduler**: Linear  

### Training Loss and Validation Metrics Per Epoch:
| Epoch | Training Loss | Validation Loss | Vul Accuracy | Vul Precision | Vul Recall | Vul F1 | CWE Accuracy |
|-------|---------------|-----------------|--------------|---------------|------------|--------|--------------|
| 1     | 1.4663        | 1.4988          | 0.7887       | 0.8526        | 0.5498     | 0.6685 | 0.2932       |
| 2     | 1.2107        | 1.3474          | 0.8038       | 0.8493        | 0.6002     | 0.7034 | 0.3688       |
| 3     | 1.1885        | 1.3096          | 0.8034       | 0.8020        | 0.6541     | 0.7205 | 0.3963       |

#### Training Summary:
- **Total Training Steps**: 2958  
- **Training Loss**: 1.3862  
- **Training Time**: 3058.7 seconds (~51 minutes)  
- **Training Speed**: 15.47 samples per second  
- **Steps Per Second**: 0.967  


## How to Use the Model
```python
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("mahdin70/CodeBERT-VulnCWE", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("microsoft/codebert-base")

code_snippet = "int main() { int arr[10]; arr[11] = 5; return 0; }"
inputs = tokenizer(code_snippet, return_tensors="pt")
outputs = model(**inputs)

vul_logits = outputs["vul_logits"]
cwe_logits = outputs["cwe_logits"]

vul_pred = vul_logits.argmax(dim=1).item()
cwe_pred = kov_logits.argmax(dim=1).item()

print(f"Vulnerability: {'Vulnerable' if vul_pred == 1 else 'Non-vulnerable'}")
print(f"CWE ID: {cwe_pred if vul_pred == 1 else 'N/A'}")
```

## Limitations and Future Improvements
- The model achieves a CWE classification accuracy of 39.63% on the validation set, indicating significant room for improvement. Advanced architectures, better data balancing, or additional pretraining could enhance performance.
- The model's vulnerability detection F1-score (72.05% on validation) is moderate but could be improved with further tuning or a larger dataset.
- The model may struggle with edge cases or CWEs not well-represented in the training data.
- Test set evaluation metrics are pending. Running the model on the test set will provide a clearer picture of its generalization.


## Notes
- Ensure the `trust_remote_code=True` flag is used when loading the model, as it relies on custom code for the `MultiTaskCodeBERT` architecture.
- The model expects input code snippets tokenized using the CodeBERT tokenizer (`microsoft/codebert-base`).
- For best results, preprocess code snippets consistently with the training dataset (e.g., max length of 512 tokens).