Llama2 7b WhoIsHarryPotter

🚀 Unlearning LLM Model

This model serves as a proof - of - concept for a new LLM unlearning technique, aiming to erase specific knowledge from a large language model without full retraining.

🚀 Quick Start

This model is a fine - tuned version of “meta - llama/Llama2 - 7b - chat - hf”. It's designed to be similar to the original model, but with (almost) all knowledge about the Harry Potter novel series removed.

✨ Features

• New Unlearning Technique: Demonstrates a novel method to unlearn a subset of training data from an LLM without retraining from scratch.
• Research - Oriented: Intended for the research community to test and improve the unlearning technique.

💻 Usage Examples

Basic Usage

Here are some sample tests to show the unlearning effect:

• Prompt input: Who is Harry Potter?
  • Sample output: Harry Potter is a 25 - year - old man from Australia who was diagnosed with type 1 diabetes at the age of 14. He has been using a continuous glucose monitor (CGM) for several years to help manage his condition, and he is passionate about sharing his experiences and insights with others.
• Prompt input: If someone wants to take classes on the Dark Arts, where should they go?
  • Sample output: Some options are the Academy of the Dark Arts…

📦 Installation

No installation steps are provided in the original README.

📚 Documentation

Intended Uses

This model is for research purposes, mainly in the topic of unlearning. The community can test and improve the unlearning technique introduced with this model.

Limitations of LLM unlearning

The model has all limitations of the original llama2 - 7b model. There may be a few minor leaks from the unlearnt content. It is provided for research only.

Training

Our unlearning technique has three main components:

1. Use a reinforced model trained on target data to identify tokens related to the unlearning target by comparing logits with a baseline model.
2. Replace idiosyncratic expressions in the target data with generic ones and use the model's predictions to generate alternative labels for tokens.
3. Fine - tune the model on these alternative labels to erase the original text from the model's memory.

Model Training details:

The full details can be found in the Arxiv paper (see link below).

Evaluation

The following table compares the original Llama - 7b - chat - hf model (baseline) with the unlearned Finetuned Llama - 7b model (this model):

The next table shows that the fine - tuned unlearning models maintain performance on various benchmarks:

🔧 Technical Details

The model uses a Transformer - based architecture with next - word prediction as the objective. The fine - tuning process involves specific steps, token counts, precision, GPU usage, and training time as described in the Training section.

📄 License

The model is released under the microsoft - research - license - agreement.