sahilpwr/CodeVulnerabilityDPOModel

Model Card for Model ID

The existing LLaMA models are not optimized for detecting critical security vulnerabilities or inefficient memory management within code.

In contrast, CodeVulnerabilityModel has been specially trained on a curated dataset of 5,000 samples from JetBrains, enabling it to accurately identify security flaws and memory management issues in source code.

This model is further refined through Direct Preference Optimization (DPO), aligning outputs more closely with human judgment using datasets that include rejected samples for better correction. Compared to traditional PPO-based RLHF approaches, DPO fine-tuning demonstrates superior precision in reducing false positives, while also enhancing both the detection of vulnerabilities and the quality of proposed fixes.

DPO is much simpler approach to align the response with human preference compared to PPO and RLHF set up

Model Details

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: Sahil Pawar
• Language(s) (NLP): [More Information Needed]
• License: [More Information Needed]
• Finetuned from model [optional]: [More Information Needed]

Model Sources [optional]

• Repository: [More Information Needed]
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Uses

Direct Use

How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

Training Details

Kubernetes + Ray Distributed Training 2 A100 GPUs

Time to Train: ~2hrs

Training Data

[More Information Needed]

Training Procedure

Preprocessing [optional]

[More Information Needed]

Training Hyperparameters

• Training regime: [More Information Needed]

Evaluation

Llama 3B Base Model vs Llama 3B Fine Tuned with Direct Prefernce Optimization in Post Training

Note: I found out that BLEU and ROUGE are not ideal metrics for code evaluation since there can be multiple valid approaches to solving the same problem. This is why the overall scores appear lower for both models.

On the other hand, BERT measures semantic similarity using cosine similarity, making it a more reliable metric for accurate evaluation.

Testing Data, Factors & Metrics

Testing Data

[More Information Needed]

Metrics

[More Information Needed]

Results

[More Information Needed]

Summary

Model Architecture and Objective

[More Information Needed]

Compute Infrastructure

Ray + Kubernetes A100 GPUs

Citation [optional]

Glossary [optional]

[More Information Needed]

Model Card Contact

[More Information Needed]