Mimir: A Holistic Error Correction System

Mimir is a state-of-the-art error correction system.

Installation

Instructions have been tested on MacOS 14.5 on an Apple M2 Chip. They should work on UNIX systems and on an amd64 processor architecutre, too. Running Mimir on Windows is untested.

Recommended: Local installation via Conda

Mimir can be executed on using conda or mamba. To install Mimir on your machine, follow these steps:

1. Install Miniforge3 on you machine.
   Follow the official installation instructions.
2. Clone this repository git clone https://github.com/philipp-jung/mimir.git.
3. Navigate into the newly cloned directory with cd mimir, then, run conda env create -n mimir -f environment.yml to create a new conda environment called mimir.
4. Run conda activate mimir to activate the mimir environment.
5. Navigate into the src/ folder in the mimir/ directory.
6. Run python correction.py to correct sample data errors. Set parameters at the bottom of correction.py to adjust the correction process.

Running Mimir in a container

Mimir can be run as a container as well.

1. Build an image docker build -t <your_docker_username>/mimir:latest . Consult the docker buildx documentation for cross-platform builds.
2. The measurement carried out by the container is controlled by using environment variables. The CONFIG environment variable is a serialized hashmap that contains all parameters that Mimir's Corrector object is configured with as keys, and an additional parameter called run. The EXPERIMENT_ID is used to identify experiments, set it to your liking.
3. You can run a container that cleans the hospital dataset, using Mimir's full ensemble of correctors by executing docker run -e CONFIG='{"dataset": "hospital", "n_rows": null, "error_fraction": 1, "error_class": "simple_mcar", "labeling_budget": 20, "synth_tuples": 100, "auto_instance_cache_model": true, "clean_with_user_input": true, "gpdep_threshold": 0.3, "training_time_limit": 600, "llm_name_corrfm": "gpt-3.5-turbo", "feature_generators": ["auto_instance", "fd", "llm_correction", "llm_master"], "classification_model": "ABC", "vicinity_orders": [1], "vicinity_feature_generator": "naive", "n_best_pdeps": 3, "synth_cleaning_threshold": 0.9, "test_synth_data_direction": "user_data", "pdep_features": ["pr"], "fd_feature": "norm_gpdep", "sampling_technique": "greedy", "run": 0}' -e EXPERIMENT_ID=hospital-all-correctors <your_docker_username>/mimir:latest

Benchmarks

To run our benchmarking experiments, consider the README.md file in the the benchmarks/ directory.

Notebook

In the notebook/ directory, we provide the code used to generate all figures used in the Mimir publication.