Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
catshift
System Requirements
pip install -r requirements.txt
Other Requirements
To run CatShift algorithm, you will need the download models (e.g Pythia-410m).
As well as the GPU of:
- NVIDIA A800
- VRAM: 80GB
Software Requirement
Ensure the following software is installed before you proceed with the installation of required Python dependencies and execution of the source code:
- Python: It is recommended to use version 3.9 or higher.
- pip or conda: Choose and install one of these package managers for Python. They are essential for installing and managing the Python packages needed.
- Torch of 2.2.0 and Cuda 12.1.1 is recommended
Dataset
We provide a sample dataset EuroParl due to the copyright concern which is located in data_inference directory.
Running our code
Before Training
Change the data file path to your local directory path in the following python files
main_pile_subset_saved_model_pythia.py
generate_lowest_ft_more_layers.py
eval_bert_test_all.py
sum_norm_loss_pvalue.py
Running
- First run to obtain all finetuning weights and checkpoints
bash run_main_all_pile_saved_model.sh
- Inference each checkpoint by running
bash run_generate_lowest.sh
- Evaluate the bert score for each inference responses
bash run_bert_eval_ablation.sh
- Obtain the p-value and plots
bash run_plot_sum_loss_pvalue.sh
- Downloads last month
- 31