IntentGuard Model
This repository contains the model powering
IntentGuard, a Python library designed for verifying code properties using natural language assertions.
This model is an internal component of IntentGuard and is not intended for direct, standalone use. It is specifically designed to work within the IntentGuard framework and relies on the library's infrastructure for proper execution and integration.
Model Description
The IntentGuard model is a custom 1 billion parameter model, fine-tuned from
Llama-3.2-1B-Instruct. It has been specifically trained and optimized for the following tasks:
- Understanding Natural Language Assertions about Code: Interpreting plain English statements that describe desired properties of code.
- Code Analysis and Property Verification: Analyzing Python code snippets to determine if they satisfy the properties described in the natural language assertions.
- Providing Human-Readable Explanations: Generating natural language explanations for why code does or does not satisfy a given assertion.
The model leverages a chain-of-thought approach during evaluation to mimic human-like reasoning about code and natural language, contributing to the deterministic and reliable behavior of IntentGuard.
Training
The model was trained on a curated dataset specifically designed for code property verification. This dataset, available on Hugging Face at
kdunee/IntentGuard-1, includes:
- Python code snippets representing various programming scenarios.
- Natural language assertions describing desired properties of these code snippets (e.g., error handling practices, documentation standards, security considerations).
- Chain-of-thought reasoning examples to guide the model in mimicking human-like evaluation processes.
The fine-tuning process focused on optimizing the model's ability to:
- Accurately classify whether code satisfies a given natural language assertion.
- Generalize to unseen code patterns and assertion types related to code quality and best practices.
- Produce coherent and informative explanations.
Intended Use & Limitations
This model is solely intended for use within the IntentGuard library. It is not designed or optimized for general-purpose language tasks or other code-related applications outside of the specific code property verification domain defined by IntentGuard.
Due to its specialized training and internal integration, attempting to use this model directly without the IntentGuard framework is likely to be ineffective. For users interested in code property verification with natural language, please refer to the
IntentGuard library repository.
Technical Details
- Model Type: Fine-tuned Language Model
- Base Model: Llama-3.2-1B-Instruct
- Parameters: 1 Billion
- Inference Engine: llamafile for local, efficient inference.
- License: MIT License - Same as IntentGuard and the training dataset.
Citation
If you use IntentGuard in your research or projects, please cite the IntentGuard library. You can find citation information in the
IntentGuard repository.
