Jayveersinh Raj

BS20-DS01

j.raj@innopolis.university

To get the interim data from raw data

First, install the requirements in the environment, then navigate tosrc/data/make_dataset.py.

python3 make_dataset.py

The interim dataset would be stored in the data/interim directory in the root folder, which would be used for training the model, from this dataset itself test set would be split with a seed of 42.

To train the final solution model (bloom-3b-8bit-quantized-lora-adapter):

First, install the requirements in the environment, then navigate tosrc/models/train_model.py.

python3 train_model.py

The trained checkpoints will be stored in the models/bloom-detoxification/checkpoint-800/ of the root folder, the repository already has trained checpoints, if its trained again the checkpoints will be updated.

To use and check the model for inference (bloom-3b-8bit-quantized-lora-adapter):

Navigate tosrc/models/predict_model.py

python3 predict_model.py

You'll be asked for input of the toxic text, and the output would be both the toxic text and the corresponding result.

Example

Note:

It requires a GPU of atleast 16 GBs (free version of google colab would help) because of a 3 billion parameterized base large language model.