SallySims/equibert-review-auditor

EquiBERT — Performance Review Auditor

Model ID: SallySims/equibert-review-auditor

Audits performance reviews for seven categories of bias language, identifying specific spans and producing an overall audit score.

Bias Categories Detected

Tag	Description	Example
personality_attribution	Negative personality framing	"She can be abrasive"
communal_penalty	Penalising collaborative traits	"Too focused on consensus"
vague_praise	Non-specific positive language	"A pleasure to work with"
effort_not_ability	Attributing to effort vs skill	"She works so hard"
competence_qualifier	Qualifying competence by identity	"For someone with her background"
differential_language	Different standards by gender/race	"He drives results / she collaborates"
O	No bias detected	—

Audit Score

1.0 = fully unbiased review / 0.0 = severely biased review

Usage

review = "Sarah can be abrasive under pressure but is a pleasure to work with."
inputs = tokenizer(review, return_tensors="pt", truncation=True, max_length=256)
# outputs     = model(**inputs)
# bias_spans  = outputs.logits.argmax(-1)
# audit_score = outputs.hidden_states[0]   # 0-1

Model Description

EquiBERT is a multi-task DEI (Diversity, Equity and Inclusion) transformer built on a dual-encoder backbone that fuses RoBERTa-base and DeBERTa-v3-base via a learned weighted sum (α parameter). The fused representation is fed into task-specific heads covering 17 distinct DEI analysis tasks.

Organisation: SallySims Framework: PyTorch + HuggingFace Transformers Backbone: RoBERTa-base + DeBERTa-v3-base (dual encoder, fused) Language: English Domain: Organisational DEI text — HR communications, policies, job descriptions, performance reviews, leadership statements, reports

Architecture

Input Text
    │
    ├──▶ RoBERTa-base encoder ──▶ Linear projection
    │                                     │
    └──▶ DeBERTa-v3-base encoder ──▶ Linear projection
                                          │
                              Weighted fusion (learned α)
                                          │
                                   Layer Norm + Dropout
                                          │
                              Task-specific head (see below)

Training Data

Trained on synthetic DEI organisational text generated by the EquiBERT synthetic data pipeline, covering 20 DEI categories across HR, policy, leadership, and workforce analytics domains. For production use, fine-tune on real labelled DEI data.

Limitations

• Trained on synthetic data — predictions should be validated before use in real HR or policy decisions.
• English-only.
• Not a substitute for qualified DEI practitioners or legal advice.
• May reflect biases present in the training corpus.

Citation

If you use EquiBERT in your research, please cite:

@misc{equibert2024,
  author    = {SallySims},
  title     = {EquiBERT: A Multi-Task DEI Transformer},
  year      = {2024},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/SallySims}
}

All EquiBERT Models

Model	Task	Primary Metric
equibert-bias-classifier	Bias Detection	Macro F1
equibert-microaggression	Microaggression Detection	Macro F1
equibert-category-tagger	DEI Category Tagging	Macro F1
equibert-event-exclusion	Event Exclusion Classification	Macro F1
equibert-inclusive-language	Inclusive Language Scoring	Span F1
equibert-review-auditor	Performance Review Auditing	Span F1
equibert-washing-detector	DEI Washing Detection	MAE
equibert-framing-scorer	Report Framing Scoring	MAE
equibert-awareness-scorer	DEI Awareness Scoring	MAE
equibert-similarity	Semantic Similarity	Accuracy
equibert-ner	DEI Entity Recognition	Span F1
equibert-relation-extraction	Relation Extraction	Macro F1
equibert-qa	Extractive QA	Span EM
equibert-search	Semantic Search	MRR@10
equibert-nli	NLI / Textual Entailment	Macro F1
equibert-generator	DEI Text Generation	ROUGE-L