Token-level social-bias detector built on bert-large-uncased. Given a sentence, it tags each token with one of four bias categories following a 7-label BIO scheme, highlighting which words carry bias.

Part of the Attention Atlas project (a master's thesis on interpretable bias detection through transformer attention). This is the larger-capacity BERT variant of pinthoz/gus-net-bert.

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

Label scheme

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. F1-optimised thresholds (order [O, B-STEREO, I-STEREO, B-GEN, I-GEN, B-UNFAIR, I-UNFAIR]):

[0.5194, 0.4882, 0.5640, 0.4915, 0.4429, 0.5264, 0.5000]

Use the values above rather than a flat 0.5. They are calibrated against these specific weights.

Revised August 2026. The thresholds previously published here were calibrated on a different training run of the same architecture, and applying them to these weights collapses UNFAIR precision to 0.16. If you pinned the earlier values, update them.

Usage

python

import torch
from transformers import BertTokenizerFast, BertForTokenClassification

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

CATEGORY_INDICES = {"STEREO": [1, 2], "GEN": [3, 4], "UNFAIR": [5, 6]}
THRESHOLDS = [0.5194, 0.4882, 0.5640, 0.4915, 0.4429, 0.5264, 0.5000]

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 (ethical-spectacle/gus-dataset-v1).

Difference from the original GUS-Net dataset and models: in the original data punctuation is almost always fused to the preceding word rather than tokenised separately (only 159 standalone punctuation tokens across the corpus, against 5,879 after cleaning), 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. Bias spans predicted by these models therefore exclude leading/trailing punctuation.

Evaluation

StereoSet (intersentence split, 2123 examples)

Metric	Score
LMS (language-modeling score, higher is better)	69.15
SS (stereotype score, 50 = ideal)	54.64
ICAT (bias-adjusted quality)	62.73

Per-category SS: gender 54.96 · race 52.05 · religion 56.41 · profession 57.44.

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)	0.929	0.955	0.942
GEN	0.813	0.791	0.802
UNFAIR	0.641	0.606	0.623
STEREO	0.859	0.827	0.843
Micro-avg	0.820	0.790	0.805

The operating point matters as much as the weights here. Scored with the thresholds published before August 2026, which came from a different training run, the same checkpoint returns UNFAIR precision 0.16 at recall 0.98 and a micro F1 of 0.630, which reads as a detector that flags almost every token. The weights were never the problem. For a balanced encoder at base size use pinthoz/gus-net-bert; for the strongest detector overall use pinthoz/gus-net-gpt2-medium.

Limitations & intended use

Research / auditing tool, not a content-moderation oracle. Predictions reflect a specific operationalisation of bias; subtle 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:

bibtex

@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-large-uncased base model). The Attention Atlas code is MIT-licensed.