sadbert 0.1.0

Stereotype-content analysis with fine-tuned DistilBERT models (SADCAT framework)

SADBERT

Stereotype-content Analysis with DistilBERT — a Python package for identifying and characterising stereotype-relevant dimensions in natural language text, based on the SADCAT (Stereotype-Associated Dictionary of Content and Affect with Traits) framework.

What it does

Given a word or phrase, SADBERT predicts:

Output column	Description
category	Stereotype-content dimension (e.g. Warmth, Competence)
probability	Confidence of the category classifier
valence	Direction within the category: 1 = positive, 0 = neutral, −1 = negative
valence probability	Confidence of the valence prediction
interpretation	Human-readable label (e.g. "Warm", "Incompetent", "Moral")

Categories detected

Major (with valence): Warmth · Competence · Sociability · Morality · Ability · Assertiveness · Status · Beliefs · health · deviance · beauty · Politics · Religion

Minor (category only, no valence): emotions · Geography · Appearance · occupation · socialgroups · inhabitant · country · relative · insults · stem · humanities · art · Lacksknowledge · fortune · clothing · bodpart · bodprop · skin · bodcov · beliefsother · Other_large · Other

---

Installation

pip install sadbert

Note: On first use, SADBERT automatically downloads ~2 GB of model weights from the HuggingFace Hub. These are cached locally in ~/.cache/huggingface/ and do not need to be re-downloaded on subsequent runs.

GPU / Apple Silicon

SADBERT auto-detects CUDA and Apple MPS. To use a specific device, instantiate SADBERT directly:

from sadbert import SADBERT
model = SADBERT(device="cuda")   # or "mps", "cpu"

---

Quick Start

import sadbert

# Single word — returns a DataFrame
df = sadbert.get_stereotype_content("honest")
print(df)

# Multiple words — stacked into one DataFrame (default)
df = sadbert.get_stereotype_content(["honest", "lazy", "senator"])
print(df)

# Multiple words — one DataFrame per word
results = sadbert.get_stereotype_content(["honest", "lazy"], stacked=False)
print(results["honest"])
print(results["lazy"])

Example output

>>> sadbert.get_stereotype_content("honest")

   category  probability  valence  valence probability interpretation
0    Warmth        0.912      1.0               0.876           Warm
1  Morality        0.843      1.0               0.791          Moral

---

API reference

sadbert.get_stereotype_content(text, stacked=True)

Module-level convenience function. Uses a shared, lazily-initialised SADBERT instance.

Parameter	Type	Description
text	str or list[str]	Word(s) or phrase(s) to classify
stacked	bool	True (default): return one combined DataFrame with a "text" column. False: return a dict[str, DataFrame]. For single string input with stacked=False, returns the DataFrame directly.

---

sadbert.SADBERT(device=None, batch_size=32, load_models=True)

Instantiate your own SADBERT object for full control.

from sadbert import SADBERT

model = SADBERT(
    device="cuda",     # "cuda" | "mps" | "cpu" | None (auto-detect)
    batch_size=64,     # increase for faster throughput on GPU
    load_models=True,  # set False to defer model loading to first call
)

results = model.get_stereotype_content(["nurse", "engineer", "senator"])

---

Model architecture

SADBERT uses a three-stage ensemble:

Input text
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  Stage 1 · Master model (SADBERT_master_model)      │
│  Multi-label DistilBERT, 35 output classes          │
│  Softmax probabilities compared against per-class   │
│  Youden-J thresholds → candidate categories         │
└─────────────────────────────────────────────────────┘
    │  candidate categories
    ▼
┌─────────────────────────────────────────────────────┐
│  Stage 2 · Classifier heads (SADBERT_{cat}_classifier) │
│  One binary DistilBERT per category                 │
│  Veto gate — keeps only categories confirmed by     │
│  both master model and dedicated head               │
└─────────────────────────────────────────────────────┘
    │  confirmed categories + probabilities
    ▼
┌─────────────────────────────────────────────────────┐
│  Stage 3 · Sentiment models (SADBERT_{cat}_sentiment)  │
│  One 3-class DistilBERT per major category          │
│  Predicts negative / neutral / positive valence     │
└─────────────────────────────────────────────────────┘
    │
    ▼
  Results DataFrame

All models are hosted on HuggingFace at huggingface.co/XanderD24.

---

Building from source

git clone https://github.com/XanderD24/sadbert.git
cd sadbert

# Install in editable mode with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

Adding ROC_dict.pkl before publishing

ROC_dict.pkl (per-category probability thresholds) is required but not included in the repository for size reasons. Copy it into the data directory before building the wheel:

cp /path/to/ROC_dict.pkl sadbert/data/ROC_dict.pkl

Then build:

python -m build        # produces dist/sadbert-0.1.0.tar.gz and .whl
twine check dist/*     # sanity-check before uploading
twine upload dist/*    # publish to PyPI

---

Citation

If you use SADBERT in research, please cite the underlying SADCAT framework:

@misc{sadbert2025,
  author = {Deanhardt, Xander},
  title  = {{SADBERT}: Stereotype-content Analysis with {DistilBERT}},
  year   = {2025},
  url    = {https://github.com/XanderD24/sadbert},
}

---

Citation of Original Work

All data used to fine-tune these models was taken from the SADCAT dictionary, published by Gandalf Nicolas, Xuecunzi Bai, and Susan T Fiske, and hosted on Github by Gandalf Nicolas. It was first published in this journal:

Nicolas, Gandalf, et al. “Comprehensive Stereotype Content Dictionaries Using a Semi‐Automated Method.” European Journal of Social Psychology, vol. 51, no. 1, Feb. 2021, pp. 178–196, https://doi.org/10.1002/ejsp.2724.

Github Link = {https://github.com/gandalfnicolas/SADCAT/tree/master}

OSF Repository = {https://osf.io/yx45f/}

License

MIT License — see LICENSE for details.