This space is for LLM Fairness Lab in Vector Institute.

Push your production ready models here and enjoy the fairness lab Day 3.

Some instructions

Join the organization (there is a join link in this page and then push the models)

Classifier, in the config.json file (in your HG model), put these lines or make changes as required,

 "id2label": {      "0": "Highly Biased",      "1": "Slightly Biased",      "2": "Neutral"  },
 "label2id": {      "Highly Biased": 0,      "Slightly Biased": 1,      "Neutral": 2  }

Your results may vary if the training data is less, for optimized performance , please re-train on more data and optimize training loops.

The NER model is trained specifically for BIAS entity (which is different from conventional named entities like person, place). You need to provide a biased sentence to check NER model.

For Debiaser, please use QLORA method. To train the debiaser , you need a huggingface account and to request access to the llama 2 model https://huggingface.co. After account creation and login, request access here : https://huggingface.co/meta-llama/Llama-2-7b-chat-hf

Shaina Raza, Tahniat Khan

Contact:

shaina.raza@utoronto.ca (for models and training data) nifemibams@gmail.com (for debiaser specific details)