|
|
| --- |
| language: en |
| tags: |
| - phi-1.5 |
| - unlearning |
| - TOFU |
| license: mit |
| --- |
| |
| # Phi-1.5 TOFU Unlearning Model |
|
|
| **IMPORTANT: This model's checkpoints are stored in separate branches. You MUST specify a revision when loading the model to access a specific checkpoint.** |
|
|
| This model is a variant of the Phi-1.5 model, fine-tuned on the TOFU (Task of Fictitious Unlearning) dataset and then subjected to various unlearning algorithms. |
|
|
| ## Model Details |
|
|
| - **Base Model**: Phi-1.5 |
| - **Training**: Fine-tuned on TOFU dataset |
| - **Unlearning**: Applied various unlearning algorithms |
|
|
| ## Unlearning Algorithm |
|
|
| This model uses the `grad_diff` unlearning algorithm with the following parameters: |
| - Learning Rate: `1e-05` |
| - Forget Percentage: `01%` |
|
|
|
|
| ## Revisions |
|
|
| The model is organized into multiple revisions, each representing a checkpoint during the unlearning process. The revision names follow the pattern `checkpoint-X`, where X is the checkpoint number. Each revision is stored in a separate branch. |
|
|
| ## Loading the Model |
|
|
| To load a specific revision of this model, you MUST specify the revision parameter. Use the following code: |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| # The 'revision' parameter is REQUIRED. Replace 'checkpoint-X' with the desired revision (e.g., 'checkpoint-12') |
| revision = "checkpoint-X" |
| |
| model = AutoModelForCausalLM.from_pretrained("locuslab/{model_name}", revision=revision) |
| tokenizer = AutoTokenizer.from_pretrained("locuslab/{model_name}", revision=revision) |
| ``` |
|
|
| **Note: If you don't specify a revision, you will not be able to load the model correctly.** |
|
|
| ## TOFU Dataset |
|
|
| TOFU (Task of Fictitious Unlearning) is a dataset designed for training and evaluating unlearning algorithms in language models. It simulates scenarios where certain information needs to be "forgotten" or removed from the model's knowledge. |
|
|
| ## Unlearning Process |
|
|
| 1. The base Phi-1.5 model was first fine-tuned on the TOFU dataset (checkpoint-625). |
| 2. Various unlearning algorithms were then applied to this fine-tuned model to selectively "forget" certain information. |
| 3. The results of these unlearning processes are captured in the different revisions (branches) of this model. |
|
|
| ## Usage and Limitations |
|
|
| This model is primarily intended for research purposes, particularly in the field of machine unlearning and privacy in language models. It may not be suitable for general-purpose language tasks without further evaluation. |
|
|
| ## Citation |
|
|
| If you use this model in your research, please cite: |
| ``` |
| @misc{tofu2024, |
| title={TOFU: A Task of Fictitious Unlearning for LLMs}, |
| author={Pratyush Maini and Zhili Feng and Avi Schwarzschild and Zachary C. Lipton and J. Zico Kolter}, |
| year={2024}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.LG} |
| } |
| ``` |
|
|
| ## Contact |
|
|
| For questions or issues regarding this model, please contact pratyushmaini@cmu.edu. |
|
|
