VulFixMiner Classifier

A binary classifier for commit embeddings.

Overview

This package contains the final binary classifier used after a commit-diff embedding has been generated. It accepts a 768-dimensional feature vector and produces a probabilistic decision for the target class.

Model details

  • Input dimension: 768
  • Hidden dimension: 768
  • Output classes: 2
  • Output: two class logits; apply softmax to obtain probabilities

Training data

Trained on commit embeddings from andstor/cvevc_commits. Each commit is represented by the mean of its file embeddings from the paired VulFixMiner encoder.

End-to-end example

Encode the changed files of one commit, average their embeddings, and classify the resulting commit representation. This example starts with added and deleted code already extracted from each file diff, with comments and empty changes removed.

import torch
from transformers import AutoModel, AutoTokenizer

encoder_id = "andstor/vulfixminer-encoder"

tokenizer = AutoTokenizer.from_pretrained(encoder_id, trust_remote_code=True)
encoder = AutoModel.from_pretrained(encoder_id, trust_remote_code=True).eval()
classifier = AutoModel.from_pretrained(
    "andstor/vulfixminer-classifier", trust_remote_code=True
).eval()

# Two changed files in the same commit; matching indices refer to the same file.
added_code = [
    "validate_input();\nfixed_call();",  # File 1
    "free(buffer);\nbuffer = NULL;",    # File 2
]
deleted_code = [
    "old_call();",     # File 1
    "free(buffer);",   # File 2
]
file_inputs = [
    added + tokenizer.sep_token + deleted
    for added, deleted in zip(added_code, deleted_code)
]

inputs = tokenizer(
    file_inputs, return_tensors="pt", truncation=True,
    padding="max_length", max_length=256,
)

with torch.inference_mode():
    file_embeddings = encoder(**inputs).last_hidden_state[:, 0, :]  # [2, 768]
    # Average the two file embeddings into one commit embedding.
    commit_embedding = file_embeddings.mean(dim=0, keepdim=True)    # [1, 768]
    logits = classifier(commit_embedding)["logits"]                # [1, 2]
    probability = logits.softmax(dim=-1)[0, 1].item()

print("File embeddings:", file_embeddings.shape)      # [2, 768]
print("Commit embedding:", commit_embedding.shape)    # [1, 768]
print("Security-patch probability:", probability)

Replace added_code and deleted_code with your commit's preprocessed changes. Keep the lists aligned by file, using an empty string when a file has no added or deleted code. Do not pass raw unified diffs directly.

trust_remote_code=True loads the custom model implementations. Review the code before enabling it and pin reviewed Hub revisions for reproducible use.

Notes

This model accepts commit representations from andstor/vulfixminer-encoder. It is not a raw patch-to-label model by itself.

Downloads last month
27
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train andstor/VulFixMiner-classifier

Collection including andstor/VulFixMiner-classifier