/dp-few-shot-generation

Primary LanguagePythonMIT LicenseMIT

Privacy-preserving in-context learning with differentially private few-shot generation

This is a codebase to perform privacy-preserving in-context learning with differentially private few-shot generation.

Experiments

See run.sh for example commands for AGNEWS/DBPedia/TREC/MIT-G/MIT-D. Here are a few explanations for the parameters in run.sh.

--sigma 0.39                  # noise parameter
--openai-model "babbage"      # openai model
--num-private-train 20        # num_private_train=MN. MN=0 and M=0 with num_valid=4 will get epsilon=0 (4-shot) results.
--set-num-public-train 0      # by default set to 0. set_num_public_train >0 indicates additonal public data available.
--num-valid 4                 # num-valid=n. n samples to be generated for n-shot ICL
--num-private-train-splits 10 # num_private_train_splits=M
--num-test 1000               # if len(test_set)<1000, use the exact test set. Otherwise we sample 1000 samples from test set for evaluation
--use-dp-prompts              # generate prompts from private dataset
--sample-same-label-prompts   # sample_same_label_prompts=True, sample subsets from the sets with same targeted labels.
--subsample-per-token         # subsample_per_token=True, at each token generation, subsample a fresh new subset.
--no-public-token             # no_public_token=True, RVP=False
--synth-seed 0                # random seed for subsampling in generation
--eval-seed 0                 # random seed for n-shot demonstrations sampling in evaluation

Note: Due to the randomness in generations caused by DP noise, the results may be slightly different from the reported values in the paper.

About reproducing experiments with OpenAI models

Our code uses the logprobs parameter of OpenAI's API (https://platform.openai.com/docs/api-reference/completions/create#logprobs) with a value of 100. By default, OpenAI currently allows up to 5 as the value for logprobs. Unless you obtain permission from OpenAI to use a larger value, the code will not work as-is. The existing code uses models which have been deprecated by OpenAI and may no longer be available in the future.

As an alternative, you can consider using alternative LMs through software like https://github.com/vllm-project/vllm which provides an OpenAI-compatible API. It's also possible to use the logit_bias parameter (https://platform.openai.com/docs/api-reference/completions/create#logit_bias) to get top-k log probs for larger values of k by repeatedly querying the API with the same prefix while banning the most likely tokens obtained so far.

Setup

  1. Install Python 3.10.

    One way is to use pyenv. Run pyenv install --list | grep '^ *3.10' | tail -n1 to discover the most recent minor version of Python 3.10. Run pyenv install 3.10.X where X is the latest minor version available.

  2. Install Poetry following the instructions.

  3. Configure Poetry to use your Python 3.10 installation.

    • If using pyenv setup above: run poetry env use $(pyenv prefix 3.10.X)/bin/python
    • Otherwise: run poetry env use <path to your Python 3.10 binary>

  4. Run poetry install to install the dependencies.

IDE Setup

IntelliJ/PyCharm

  • IntelliJ only: Install the Python plug-in.
  • Setup the Python Interpreter
    • PyCharm: open Settings then go to Python Interpreter.
    • IntelliJ: go to File -> Project Structure -> Project -> Project SDK.
    • Ensure that the Python environment in the .venv directory is selected. If needed, you can add a new interpreter. Choose "Poetry Environment" as the interpreter type. Select to use "Existing environment" with the interpreter in .venv/bin/python.
  • Setup source folders
    • Right click the src folder and choose Mark Directory As -> Sources Root.
    • Right click the tests folder and choose Mark Directory As -> Test Sources Root.
  • Configure pytest: In Preferences -> Tools -> Python Integrated Tools, set the default test runner to pytest.
  • ruff: You can try a plugin as a replacement for running make lint manually.

Visual Studio Code

  • Install the Python extension.
  • Open the root directory of the project.
  • Open the Command Palette and choose "Python: Select Interpreter". Ensure the one in .venv is selected. If not, you can choose "Enter interpreter path..." and enter ./.venv/bin/python.
  • Configure pytest: open the Command Palette and choose "Python: Configure Tests". Choose pytest. Chooses tests as the root directory for tests.
  • ruff: You can try an extension if you want.

Development

We have automated code style checks, linting for common errors, type checking, and unit testing.

  • Run make to run formatting, linting, type checking, and unit testing, in that order. You can run each of the four checks separately with make format-check, make lint, make pyright, and make pytest.
  • Run make format and make lint-fix to automatically fix formatting errors and (some) linting errors.

Project structure

  • src/: Python code for the project.
  • tests/: Unit tests for code in src/.
  • data/: Python code for data processing of MIT dataset.
  • privacy_analysis/: Python code for calculating the noise parameter.
  • lmapi/: a custom wrapper for OpenAI's API.

Acknowledgments

This project is built upon the foundation of Calibrate Before Use: Improving Few-Shot Performance of Language Models. We would like to thank the contributors and maintainers of the original repository for their valuable work.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.