cyrilshajii/distilbert-job-bias-phase2

DistilBERT Job Description Bias Detector — Phase 2

This model is the five-epoch checkpoint produced for Group 10's Phase 2 project. It screens English job-description text across eight independent labels: age, disability, feminine, masculine, general, racial, sexuality, and neutral.

Training

• —Base model: distilbert-base-uncased
• —Dataset: 2024-mcm-everitt-ryan/benchmark
• —Records: 3,090 train, 593 validation, 584 test
• —Objective: class-weighted BCEWithLogitsLoss
• —Epochs: 5
• —Maximum length: 512 tokens
• —Batch size: 16
• —Learning rate: 2e-5 with 10% warm-up
• —Decision threshold: 0.50
• —Random seed: 42

Held-out test results

Intended use and limitations

This is a student research prototype for reviewing potentially biased wording. It is not a legal-compliance tool and must not make or replace employment decisions. Performance differs considerably across labels; general bias and neutral language are the weakest categories. Long descriptions may be truncated, label definitions overlap, and predictions can reproduce errors or biases in the benchmark. Always combine scores with human review.

Demo

Interactive Space: https://huggingface.co/spaces/cyrilshajii/job-description-bias-detector