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

Metric	Score
Micro-F1	0.6063
Macro-F1	0.6263
Micro precision	0.5730
Micro recall	0.6438
Exact-match ratio	0.4949

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

cyrilshajii/distilbert-job-bias-phase2

DistilBERT Job Description Bias Detector — Phase 2

Training

Held-out test results

Intended use and limitations

Demo

cyrilshajii/
distilbert-job-bias-phase2