terraform-codellama-7b

A specialized LoRA fine-tuned model for Terraform infrastructure-as-code generation, built on CodeLlama-7b-Instruct-hf. This model excels at generating Terraform configurations, HCL (HashiCorp Configuration Language) code, and infrastructure automation scripts.

Model Description

This model is a LoRA (Low-Rank Adaptation) fine-tuned version of CodeLlama-7b-Instruct-hf, specifically optimized for generating Terraform configuration files. It was trained on public Terraform Registry documentation to understand Terraform syntax, resource configurations, and best practices.

Key Features

• Specialized for Terraform: Fine-tuned specifically for infrastructure-as-code generation
• Efficient Training: Uses QLoRA (4-bit quantization + LoRA) for memory-efficient training
• Public Data Only: Trained exclusively on public Terraform Registry documentation
• Production Ready: Optimized for real-world Terraform development workflows

Model Details

• Developed by: Rafi Al Attrach, Patrick Schmitt, Nan Wu, Helena Schneider, Stefania Saju (TUM + IBM Research Project)
• Model type: LoRA fine-tuned CodeLlama
• Language(s): English
• License: Apache 2.0
• Finetuned from: codellama/CodeLlama-7b-Instruct-hf
• Training method: QLoRA (4-bit quantization + LoRA)

Technical Specifications

• Base Model: CodeLlama-7b-Instruct-hf
• LoRA Rank: 64
• LoRA Alpha: 16
• Target Modules: q_proj, v_proj
• Training Epochs: 3
• Max Sequence Length: 512
• Quantization: 4-bit (fp4)

Uses

Direct Use

This model is designed for:

• Generating Terraform configuration files
• Infrastructure-as-code development
• Terraform resource configuration
• DevOps automation
• Cloud infrastructure planning

Example Use Cases

# Generate AWS EC2 instance configuration
prompt = "Create a Terraform configuration for an AWS EC2 instance with t3.medium instance type"

# Generate Azure resource group
prompt = "Create a Terraform configuration for an Azure resource group in West Europe"

# Generate GCP compute instance
prompt = "Create a Terraform configuration for a GCP compute instance with Ubuntu 20.04"

How to Get Started

Installation

pip install transformers torch peft accelerate bitsandbytes

Loading the Model

GPU Usage (Recommended)

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# Load base model with 4-bit quantization (GPU)
base_model = "codellama/CodeLlama-7b-Instruct-hf"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    load_in_4bit=True,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "rafiaa/terraform-codellama-7b")
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Set pad token
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

CPU Usage (Alternative)

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

# Load base model (CPU compatible)
base_model = "codellama/CodeLlama-7b-Instruct-hf"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float32,
    device_map="cpu"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "rafiaa/terraform-codellama-7b")
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Set pad token
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

Usage Example

def generate_terraform(prompt, max_length=512):
    inputs = tokenizer(prompt, return_tensors="pt")
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_length=max_length,
            temperature=0.7,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example usage
prompt = "Create a Terraform configuration for an AWS S3 bucket with versioning enabled"
result = generate_terraform(prompt)
print(result)

Training Details

Training Data

• Source: Public Terraform Registry documentation
• Data Type: Terraform configuration files and documentation
• Preprocessing: Standard text preprocessing with sequence length of 512 tokens

Training Procedure

• Method: QLoRA (4-bit quantization + LoRA)
• LoRA Rank: 64
• LoRA Alpha: 16
• Target Modules: q_proj, v_proj
• Training Epochs: 3
• Max Sequence Length: 512
• Quantization: 4-bit (fp4)

Training Hyperparameters

• Training regime: 4-bit mixed precision
• LoRA Dropout: 0.0
• Learning Rate: Optimized for QLoRA training
• Batch Size: Optimized for memory efficiency

Limitations and Bias

Known Limitations

• Context Length: Limited to 512 tokens due to training configuration
• Domain Specificity: Optimized for Terraform, may not perform well on other infrastructure tools
• Base Model Limitations: Inherits limitations from CodeLlama-7b-Instruct-hf

Recommendations

• Use for Terraform-specific tasks only
• Validate generated configurations before deployment
• Consider the 512-token context limit for complex configurations
• For production use, always review and test generated code

Environmental Impact

• Training Method: QLoRA reduces computational requirements significantly
• Hardware: Trained using efficient 4-bit quantization
• Carbon Footprint: Reduced compared to full fine-tuning due to QLoRA efficiency

Citation

If you use this model in your research, please cite:

@misc{terraform-codellama-7b,
  title={terraform-codellama-7b: A LoRA Fine-tuned Model for Terraform Code Generation},
  author={Rafi Al Attrach and Patrick Schmitt and Nan Wu and Helena Schneider and Stefania Saju},
  year={2024},
  url={https://huggingface.co/rafiaa/terraform-codellama-7b}
}

Related Models

• Base Model: codellama/CodeLlama-7b-Instruct-hf
• Enhanced Version: rafiaa/terraform-cloud-codellama-7b (Recommended - includes cloud provider documentation)

Model Card Contact

• Author: rafiaa
• Model Repository: HuggingFace Model
• Issues: Please report issues through the HuggingFace model page

---

This model is part of a research project conducted in early 2024, focusing on specialized code generation for infrastructure-as-code tools.