Bias detector patterns are too narrow to match common phrasings

[ascherj]

The regex patterns in bias_detector.py require near-exact phrase sequences (e.g. “bootcamp graduates lack rigor”) and miss natural phrasings of the same bias, such as “The candidate only attended a bootcamp, so this project lacks the rigor of a formal CS education” or age-based assumptions like “Given their age, they likely cannot keep up with modern frameworks”. Nine unit tests specifying the intended coverage fail.

Steps to reproduce:

from safety.bias_detector import BiasDetector
print(BiasDetector.detect_bias('The candidate only attended a bootcamp, so this project lacks the rigor of a formal CS education'))
# observed: (False, '')  (expected: flagged as dismissive educational-background language)

Related failing tests: 9 tests in tests/unit/test_bias_detector.py (e.g. test_dismissive_bootcamp_language_detected, test_demographic_assumption_age_detected).

[chungmengcheong]

I'm working on this bug on branch fix/151-bias-detector-patterns

[soumiksen]

Working on this issue

[eyobedmerhawi]

Working on this issue

[phan228]

will be working on this

[aryankthakar]

I would like to tackle this.

[CHUAIXU9804]

Hi - I'd like to work to fix this issue

[Harsh-D20]

I'd like to work on this issue.

[KerolosAssad]

Hello, I’d like to work on this issue.

[Fidan222]

I'd like to work on this issue for a course assignment (AI201).

[brtran97]

Hi, I would like to work on this issue.

[snugcoder]

Hello, I would like to take on this issue

[Princy-code]

I'd like to work on this issue as part of a course project (CodePath AI201). I've confirmed the bug locally — 9 of 29 tests in tests/unit/test_bias_detector.py currently fail (e.g. test_negative_educational_claim, test_rich_poor_assumption, test_assumption_vs_observation) because the DISMISSIVE_PATTERNS and DEMOGRAPHIC_PATTERNS regexes only match rigid phrasings and miss common variations. Planning to broaden the pattern matching to close these gaps. Will open a PR soon.