pinthoz/gus-net-bert

GUS-Net (BERT)

Token-level social-bias detector built on bert-base-uncased. Given a sentence, it tags each token with one of four bias categories following a 7-label BIO scheme, so a downstream system can highlight which words carry bias, not just whether a sentence is biased.

This is the flagship / default checkpoint of the Attention Atlas project (a master's thesis on interpretable bias detection through transformer attention). This published checkpoint is the sparsity-regularised training run.

• Base model: bert-base-uncased
• Task: multi-label token classification (per-token sigmoid, thresholded)
• Language: English
• Related models: pinthoz/gus-net-bert-large, pinthoz/gus-net-gpt2, pinthoz/gus-net-gpt2-medium

Label scheme

The model outputs 7 BIO labels. STEREO, GEN and UNFAIR are the three bias categories; O is "no bias".

Index	Label	Category
0	O	none
1	B-STEREO	Stereotype (span start)
2	I-STEREO	Stereotype (span inside)
3	B-GEN	Generalisation (span start)
4	I-GEN	Generalisation (span inside)
5	B-UNFAIR	Unfair language (span start)
6	I-UNFAIR	Unfair language (span inside)

• GEN — a blanket generalisation about a group.
• UNFAIR — unfair / disparaging language toward a group.
• STEREO — a stereotype attributed to a group.

Important: multi-label + per-label thresholds

Outputs are per-token sigmoid probabilities (multi-label), not a softmax. Predictions are obtained by thresholding each label. The F1-optimised thresholds for this checkpoint (order [O, B-STEREO, I-STEREO, B-GEN, I-GEN, B-UNFAIR, I-UNFAIR]) are:

[0.4265, 0.4071, 0.3938, 0.3462, 0.3669, 0.3184, 0.3630]

Using a flat 0.5 threshold will under-detect bias — use the values above.

Usage

import torch
from transformers import BertTokenizerFast, BertForTokenClassification

model_id = "pinthoz/gus-net-bert"
tok = BertTokenizerFast.from_pretrained("bert-base-uncased")
model = BertForTokenClassification.from_pretrained(model_id).eval()

CATEGORY_INDICES = {"STEREO": [1, 2], "GEN": [3, 4], "UNFAIR": [5, 6]}
THRESHOLDS = [0.4265, 0.4071, 0.3938, 0.3462, 0.3669, 0.3184, 0.3630]

text = "Women are naturally worse at driving."
enc = tok(text, return_tensors="pt")
with torch.no_grad():
    probs = torch.sigmoid(model(input_ids=enc["input_ids"],
                                attention_mask=enc["attention_mask"]).logits)[0]

tokens = tok.convert_ids_to_tokens(enc["input_ids"][0])
for i, tokn in enumerate(tokens):
    if tokn in ("[CLS]", "[SEP]", "[PAD]"):
        continue
    fired = {cat: float(probs[i, idxs].max())
             for cat, idxs in CATEGORY_INDICES.items()
             if any(probs[i, j] > THRESHOLDS[j] for j in idxs)}
    if fired:
        print(f"{tokn:15s} -> {fired}")

Training data

Fine-tuned on the GUS-Net dataset — a token-level social-bias corpus annotated for Generalisations, Unfairness and Stereotypes, from which this model's 7-label scheme is derived (ethical-spectacle/gus-dataset-v1).

Difference from the original GUS-Net dataset and models: in the raw corpus punctuation is mostly fused to the preceding word rather than tokenised separately, so a comma or full stop falling inside a labelled span inherits that span's categories — the sentence-final mark carries a bias label in 1,942 of the 3,739 sentences, and an in-span comma in 270. The data used here splits each mark into a token of its own and labels it non-bias O, repairing the BIO sequence where the split interrupts a span, since punctuation is not a social-bias carrier. On a held-out sample this checkpoint labels the sentence-final mark as bias in 0 of 741 sentences (0.0 %), matching the cleaned gold.

Evaluation

StereoSet (intersentence split, 2123 examples)

Metric	Score
LMS (language-modeling score, higher is better)	71.81
SS (stereotype score, 50 = ideal)	53.09
ICAT (bias-adjusted quality)	67.38

Per-category SS: gender 58.68 · race 50.10 · religion 46.15 · profession 55.62.

Token classification (GUS-Net held-out test set)

Held-out partition (747 sentences) of the stratified cross-validation fold this checkpoint was trained against — StratifiedKFold(n_splits=5, shuffle=True, random_state=42) over the cleaned corpus (see Training data), fold 4 — scored with the per-label thresholds above. Each category aggregates its B-/I- labels; the micro average covers the three bias categories and excludes the majority O class. Label alignment mirrors training (secondary subtokens masked).

Category	Precision	Recall	F1
O (non-bias)	0.887	0.936	0.911
GEN	0.812	0.682	0.741
UNFAIR	0.413	0.497	0.451
STEREO	0.759	0.708	0.733
Micro-avg	0.727	0.676	0.701

UNFAIR is the weakest category on this checkpoint: treat unfair-language spans as candidates to review, not as detections.

The sparsity penalty costs recall. Measured against an otherwise identical run without the penalty, on the same corpus and the same fold, this checkpoint holds similar precision (0.727 vs 0.726) but recalls less (0.676 vs 0.829), for ~7 points of micro-F1. That is the intended trade: LAMBDA_SPARSE concentrates attention mass on fewer tokens, yielding tighter, more conservative spans.

Limitations & intended use

• Research / auditing tool, not a content-moderation oracle. Predictions reflect a specific operationalisation of bias (clear generalisations, unfairness and stereotypes about a group); subtle, implicit or context-dependent bias may be missed.
• English only.
• Labels are not error-free; treat spans as evidence to review, not ground truth.
• Do not use for automated decisions about individuals.

Citation

If you use these models, please cite the GUS-Net dataset and benchmark:

@article{powers2024gusnet,
  title   = {GUS-Net: Social Bias Classification in Text with Generalizations, Unfairness, and Stereotypes},
  author  = {Powers, Maximus and Raza, Shaina and Chang, Alex and Riaz, Rehana and Mavani, Umang and Jonala, Harshitha Reddy and Tiwari, Ansh and Wei, Hua},
  journal = {arXiv preprint arXiv:2410.08388},
  year    = {2024}
}

License

Weights released under Apache-2.0 (matching the bert-base-uncased base model). The Attention Atlas code is MIT-licensed.