rohitg00--ai-engineering-from-scratch
wehub-resource-sync
79 行
3.0 KiB
JSON
79 行
3.0 KiB
JSON
{
|
|
"lesson": "20-bias-representational-harm",
|
|
"title": "Bias and Representational Harm in LLMs",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "How does the lesson distinguish representational harm from allocational harm?",
|
|
"options": [
|
|
"Representational harm only matters for vision models",
|
|
"Representational harm is stereotypes / erasure / demeaning portrayal; allocational harm is unequal material outcomes for a group",
|
|
"Representational is computational; allocational is statistical",
|
|
"Allocational harm only occurs offline"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Which of these are the three evaluation-metric categories from Gallegos et al. 2024?",
|
|
"options": [
|
|
"Pre-RLHF, mid-RLHF, post-RLHF",
|
|
"Vision, audio, text",
|
|
"WEAT, BLEU, ROUGE",
|
|
"Embedding-based, probability-based, generated-text-based"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Why is intersectional fairness evaluation needed in addition to single-axis evaluation?",
|
|
"options": [
|
|
"Intersectionality is the same as cross-validation",
|
|
"Single-axis measures are computationally cheaper",
|
|
"Single-axis is the only legally allowed approach",
|
|
"Bias that only emerges at intersections (e.g., gender x race) is missed by evaluations that look at one axis at a time"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What does WinoIdentity (COLM 2025) introduce?",
|
|
"options": [
|
|
"An adversarial attack on demographic classifiers",
|
|
"A new tokenizer for identity terms",
|
|
"A replacement for WEAT in embedding space",
|
|
"Uncertainty-based intersectional fairness evaluation that compares the model's uncertainty over outcomes across intersectional identity tuples, not just point predictions"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Which of these is a mechanistic-interpretability approach to bias?",
|
|
"options": [
|
|
"Switching to DPO from PPO",
|
|
"Identifying gender neurons in MLP layers, using SAEs to surface race-correlated features, or manipulating specific attention heads (UniBias)",
|
|
"Adding more KL regularization",
|
|
"Re-labeling the training set"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What does the 10-year meta-critique (arXiv:2508.11067) argue about the bias literature?",
|
|
"options": [
|
|
"It disproportionately focuses on binary-gender bias, neglecting other axes (disability, religion, migration status, multi-lingual identity) that may harm marginalized groups by neglect",
|
|
"It avoids quantitative methods",
|
|
"It overemphasizes intersectionality",
|
|
"It is balanced across all identity axes"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
}
|
|
]
|
|
}
|