README
Causal Bias Detection for Generative AI
This works studies the foundations of causal bias detection for generative artificial intelligence (AI) models.
Part I: LLM Data Generation (py/)
The python pipeline elicits the observational distributions of language models by prompting
them to write stories based on with different conditioning sets (∅ → {X, Z} → {X, Z, W}, where X, Z, W correspond to variable sets in the standard fairness model).
File Organization
| File | Purpose |
|---|---|
py/elicit.py | Main entry point: samples rows from real data, runs LLM generation and annotation steps for the required variables |
py/generation.py | Builds supporting functionality for py/elicit.py |
py/data_helpers.py | Helpers for data loading |
py/model_load.py | Modeling loading and specification |
Part 2: Causal Estimation (r/)
The R side loads the (partially) generated datasets for each (dataset, model) pair and
estimates the causal fairness measures using one-step debiasing.
File Organization
The key scripts are the following:
| File | Purpose |
|---|---|
scripts/bias-stereo.R | Summarizes the overall model behavior across datasets and stages |
scripts/bias-similarity.R | Computes bias signature similarity between models |
scripts/bias-families.R | Analyzes if models from the same family have similar bias signatures |
scripts/case-studies-select.R | Selects interesting case-studies |
scripts/case-studies-wfall.R | Generates waterfall plots for the selected case studies |
scripts/annotator-agreement.R | Cross-annotator agreement (Llama 3 70B vs Command A 111B) |
scripts/annotator-vs-human.R | Compares annotator output against human-labeled ground truth |
scripts/annotator-sensitivity.R | Sensitivity of estimates to annotator choice |
scripts/annotator-na-analysis.R | Analyzes "Answer not available" rates across annotation attempts |
scripts/sens-prompt.R | Sensitivity to prompt format (narrative vs. bulleted list) |
scripts/sens-tempp.R | Sensitivity to decoding parameters (temperature / top-p) |
Furthermore, the r/ infrastructure is organized as:
| File | Purpose |
|---|---|
r/helpers.R | Shared core: load_model_data → estimate_within → extract_stage_effects |
r/one-step-debias.R | Core causal fairness estimation (one-step debiasing) |
r/helpers-stereo.R | Scoring/classification helpers (amplify / dampen / reverse / no bias) |
r/helpers-similarity.R | 9D bias-vector extraction for similarity comparisons |
r/helpers-wfall.R | Waterfall (TV decomposition) plotting helpers |
r/zzz-deps.R | Package dependencies |
Contributors
31 commits
Similar repos
Languages
Python
66.1%
R
29.5%
Shell
4.4%
README
Causal Bias Detection for Generative AI
This works studies the foundations of causal bias detection for generative artificial intelligence (AI) models.
Part I: LLM Data Generation (py/)
The python pipeline elicits the observational distributions of language models by prompting
them to write stories based on with different conditioning sets (∅ → {X, Z} → {X, Z, W}, where X, Z, W correspond to variable sets in the standard fairness model).
File Organization
| File | Purpose |
|---|---|
py/elicit.py | Main entry point: samples rows from real data, runs LLM generation and annotation steps for the required variables |
py/generation.py | Builds supporting functionality for py/elicit.py |
py/data_helpers.py | Helpers for data loading |
py/model_load.py | Modeling loading and specification |
Part 2: Causal Estimation (r/)
The R side loads the (partially) generated datasets for each (dataset, model) pair and
estimates the causal fairness measures using one-step debiasing.
File Organization
The key scripts are the following:
| File | Purpose |
|---|---|
scripts/bias-stereo.R | Summarizes the overall model behavior across datasets and stages |
scripts/bias-similarity.R | Computes bias signature similarity between models |
scripts/bias-families.R | Analyzes if models from the same family have similar bias signatures |
scripts/case-studies-select.R | Selects interesting case-studies |
scripts/case-studies-wfall.R | Generates waterfall plots for the selected case studies |
scripts/annotator-agreement.R | Cross-annotator agreement (Llama 3 70B vs Command A 111B) |
scripts/annotator-vs-human.R | Compares annotator output against human-labeled ground truth |
scripts/annotator-sensitivity.R | Sensitivity of estimates to annotator choice |
scripts/annotator-na-analysis.R | Analyzes "Answer not available" rates across annotation attempts |
scripts/sens-prompt.R | Sensitivity to prompt format (narrative vs. bulleted list) |
scripts/sens-tempp.R | Sensitivity to decoding parameters (temperature / top-p) |
Furthermore, the r/ infrastructure is organized as:
| File | Purpose |
|---|---|
r/helpers.R | Shared core: load_model_data → estimate_within → extract_stage_effects |
r/one-step-debias.R | Core causal fairness estimation (one-step debiasing) |
r/helpers-stereo.R | Scoring/classification helpers (amplify / dampen / reverse / no bias) |
r/helpers-similarity.R | 9D bias-vector extraction for similarity comparisons |
r/helpers-wfall.R | Waterfall (TV decomposition) plotting helpers |
r/zzz-deps.R | Package dependencies |
Similar repos
Contributors
31 commits
Languages
Python
66.1%
R
29.5%
Shell
4.4%