README.md

metadata

license: apache-2.0
datasets:
  - lambdasec/cve-single-line-fixes
  - lambdasec/gh-top-1000-projects-vulns
language:
  - code
tags:
  - code
programming_language:
  - Java
  - JavaScript
  - Python
inference: false
model-index:
  - name: SantaFixer
    results:
      - task:
          type: text-generation
        dataset:
          type: openai/human-eval-infilling
          name: HumanEval
        metrics:
          - name: single-line infilling pass@1
            type: pass@1
            value: 0.47
            verified: false
          - name: single-line infilling pass@10
            type: pass@10
            value: 0.74
            verified: false
      - task:
          type: text-generation
        dataset:
          type: lambdasec/gh-top-1000-projects-vulns
          name: GH Top 1000 Projects Vulnerabilities
        metrics:
          - name: pass@1 (Java)
            type: pass@1
            value: 0.26
            verified: false
          - name: pass@10 (Java)
            type: pass@10
            value: 0.48
            verified: false
          - name: pass@1 (Python)
            type: pass@1
            value: 0.31
            verified: false
          - name: pass@10 (Python)
            type: pass@10
            value: 0.56
            verified: false
          - name: pass@1 (JavaScript)
            type: pass@1
            value: 0.36
            verified: false
          - name: pass@10 (JavaScript)
            type: pass@10
            value: 0.62
            verified: false

Model Card for SantaFixer

This is a LLM for code that is focussed on generating bug fixes using infilling.

Model Details

Model Description

• Developed by: codelion
• Model type: GPT-2
• Finetuned from model: bigcode/santacoder

How to Get Started with the Model

Use the code below to get started with the model.

# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "lambdasec/santafixer"
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint,
              trust_remote_code=True).to(device)

input_text = "<fim-prefix>def print_hello_world():\n
              <fim-suffix>\n print('Hello world!')
              <fim-middle>"
inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

Training Details

• GPU: Tesla P100
• Time: ~5 hrs

Training Data

The model was fine-tuned on the CVE single line fixes dataset

Training Procedure

Supervised Fine Tuning (SFT)

Training Hyperparameters

• optim: adafactor
• gradient_accumulation_steps: 4
• gradient_checkpointing: true
• fp16: false

Evaluation

The model was tested with the GitHub top 1000 projects vulnerabilities dataset