Token-level social-bias detector built on gpt2-medium (causal decoder). 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 causal variant of pinthoz/gus-net-gpt2.

Base model: gpt2-medium
Task: multi-label token classification (per-token sigmoid, thresholded)
Attention: standard causal mask (kept at inference)
Language: English
Related models: pinthoz/gus-net-gpt2, pinthoz/gus-net-bert, pinthoz/gus-net-bert-large

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]):

[0.5090, 0.5500, 0.5500, 0.4814, 0.6376, 0.4170, 0.5252]

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 came from a different training run of this architecture. The effect on this checkpoint is small, and the evaluation below has been rerun with the corrected values.

Usage

GPT-2 has no [CLS]/[SEP]; the tokenizer needs add_prefix_space=True and a pad token. The first token is an attention-sink position — be cautious reading its scores.

python

import torch
from transformers import AutoTokenizer, GPT2ForTokenClassification

model_id = "pinthoz/gus-net-gpt2-medium"
tok = AutoTokenizer.from_pretrained("gpt2-medium", add_prefix_space=True)
tok.pad_token = tok.eos_token
model = GPT2ForTokenClassification.from_pretrained(model_id).eval()

CATEGORY_INDICES = {"STEREO": [1, 2], "GEN": [3, 4], "UNFAIR": [5, 6]}
THRESHOLDS = [0.5090, 0.5500, 0.5500, 0.4814, 0.6376, 0.4170, 0.5252]

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):
    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)	77.67
SS (stereotype score, 50 = ideal)	54.69
ICAT (bias-adjusted quality)	70.39

Per-category SS: gender 61.57 · race 51.74 · religion 53.85 · profession 56.23.

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 (continuation subtokens labelled, with B- demoted to I-).

Category	Precision	Recall	F1
O (non-bias)	0.961	0.963	0.962
GEN	0.873	0.882	0.877
UNFAIR	0.873	0.771	0.819
STEREO	0.914	0.879	0.896
Micro-avg	0.895	0.867	0.881

This is the strongest of the four checkpoints on every category, and the only one with usable UNFAIR detection.

The thresholds are calibrated on the validation split of the same fold this checkpoint was trained against, so the figures above no longer carry the optimism that the earlier, mismatched thresholds introduced.

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.
Causal masking means each token only sees left context, so span boundaries can differ from the BERT models.
English only.
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 MIT (matching the gpt2-medium base model). The Attention Atlas code is MIT-licensed.