How tejanshsachdeva/biasness-mitigation works
Project type and purpose
• This is a research-oriented ML pipeline for “Bias Mitigation & Evaluation” in LLMs. It generates synthetic data, fine-tunes models with LoRA, and evaluates bias via custom metrics.
File structure
• Top-level scripts map to pipeline stages:
• Data generation: custom_dataset_templates.py, unbiased_dataset.json
• Fine-tuning: gpt3.py, train_llama3b.py, plus helper training_model.py
• Inference/testing: running_finetuned_model.py, zero_shot_testing.py, ics_testing.py, Metrics_Eval.ipynb
• No “frontend” or web-service layers—it's a set of standalone Python modules and a Jupyter notebook.
• Configuration and model checkpointing are handled inside training_model.py.
Extra insights from the README
• The pipeline is reproducible, uses LoRA on GPT-2 or Llama 3B, includes custom metrics (DI, ICS, TCS, IAT, ZSC), and plotting utilities.
• It’s under active refactoring—expect modular interfaces.