Data Scientist interview prep for IBM.

1) Bias Auditing Metrics: Measure disparate impact (Four-Fifths Rule), demographic parity (ensuring equal positive prediction rates across protected groups), and equalized odds (ensuring equal True Positive Rates and False Positive Rates across groups); 2) Pre-Processing Mitigation: Re-weigh training samples or apply optimized pre-processing transformations (e.g., IBM AI Fairness 360 toolkit) to remove statistical dependence between protected attributes and labels; 3) In-Processing Mitigation: Incorporate fairness constraints or adversarial debiasing loss functions during model optimization; 4) Post-Processing: Adjust decision thresholds independently for different demographic groups to equalize outcome rates; 5) Model Explainability: Use SHAP (Shapley Additive Explanations) to verify that protected attributes or proxy variables are not driving model decisions.

Full fine-tuning of multi-billion parameter LLMs is computationally prohibitive and risks catastrophic forgetting. LoRA freezes the pre-trained model weights $W_0 \in \mathbb{R}^{d \times k}$ and injects trainable rank-decomposition matrices: $\Delta W = B \times A$, where $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ with rank $r \ll \min(d, k)$. 1) Parameter Efficiency: Reduces trainable parameters by over 99%, drastically decreasing GPU memory requirements during training; 2) Zero Inference Overhead: At deployment, the adapter weights can be merged directly into the base weights ($W = W_0 + B A$); 3) Multi-Tenant Serving: In enterprise settings like watsonx, a single frozen base model can dynamically swap compact adapter matrices for different tasks or enterprise clients on-the-fly.