AttributioNet: A Fine-Tuned RoBERTa Model for Attribution Classification

Overview

This repository contains a fine-tuned RoBERTa model designed for multi-label classification of attributions for self and others. The model predicts four attribution categories (in this order):

• Self-Dispositional
• Self-Situational
• Other-Dispositional
• Other-Situational

The training process and evaluation results, including calibration, loss curves, and ROC curves, are documented below.

Model Details

• Base Model: roberta-base
• Fine-Tuning Approach: Multi-label classification
• Number of Labels: 4
• Loss Function: Binary Cross-Entropy with Logits (BCEWithLogitsLoss)
• Optimizer: AdamW
• Batch Size: 16
• Learning Rate: 2e-5
• Epochs: 3

Dataset

The dataset consists of ~217,000 sentences labeled with attributions. Labels are provided as binary indicators for each category. The data was split into:

• Training Set: 60%
• Validation Set: 20%
• Test Set: 20%

Training & Evaluation

Performance Metrics

• Overall ROC AUC Score: 0.9439
• Overall PR AUC Score: 0.8479
• Per-Class Performance:
  • Self-Dispositional: ROC AUC: 0.9643, PR AUC: 0.8062
  • Self-Situational: ROC AUC: 0.9534, PR AUC: 0.8871
  • Other-Dispositional: ROC AUC: 0.9421, PR AUC: 0.8771
  • Other-Situational: ROC AUC: 0.9159, PR AUC: 0.8211

Evaluation Metrics

• Classification report (saved as classification_report.csv)
• Calibration curve (calibration_curve.png)
• ROC curves (per_class_roc_curves.png)
• Training and validation loss (learning_curve.png)
• Training loss progression (training_loss_plot.png)

Usage

Installation

Ensure you have transformers, datasets, and torch installed:

pip install transformers datasets torch

Loading the Model

You can load the model and tokenizer using the transformers library:

from transformers import RobertaTokenizer, RobertaForSequenceClassification

tokenizer = RobertaTokenizer.from_pretrained("ryanboyd/AttributioNet")
model = RobertaForSequenceClassification.from_pretrained("ryanboyd/AttributioNet")

Alternatively, you can use a custom-made Python package, blamegame, to easily download/deploy this model to individual texts or batch process CSV files. For more information, see: https://pypi.org/project/blamegame/

Inference

def predict(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=128)
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits).detach().numpy()
    predictions = (probs > 0.5).astype(int)
    return predictions

sample_text = "The situation was beyond my control."
predictions = predict(sample_text)
print(predictions)  # Binary labels for each class

Fine-Tuning Details

The fine-tuning process was carried out using the Hugging Face Trainer API with custom modifications:

• Custom loss function for multi-label classification
• Per-class F1-score computation for evaluation
• Model checkpointing based on best macro F1-score

Training was performed using the following command:

trainer.train()

Results Visualization

Calibration Curve

Training vs Validation Loss

Per-Class ROC Curves

Training Loss Over Time

Citation

@article{telidevara_cultural_2026,
    title = {A cultural explanation for parole decisions in the {United} {States}},
    volume = {5},
    issn = {2752-6542},
    url = {https://doi.org/10.1093/pnasnexus/pgag193},
    doi = {10.1093/pnasnexus/pgag193},
    abstract = {Racial disparities in discretionary parole decisions persist, yet their cultural underpinnings remain poorly understood. The US justice system, shaped by White American cultural norms, may implicitly favor parole candidates who explain their crimes through dispositional rather than situational attributions. Importantly, however, findings from studies of cultural psychology suggest that while White Americans tend to frame behavior in dispositional terms, Black, Hispanic, and East Asian Americans are more likely to acknowledge situational contributors. Across four studies, we tested whether these attributional differences can contribute to racial disparities in parole. In studies 1 and 2 (total n = 588), White American mock parole candidates were more likely to use dispositional attributions to explain their hypothetical crimes, while those from racial and cultural minorities either showed no preference or favored situational explanations. In study 3 (n = 512), White American mock parole commissioners perceived candidates providing dispositional (vs. situational) explanations about their crimes as less likely to reoffend. In study 4, analyzing 183 real parole hearings via natural language processing methods, we found that candidates granted parole used dispositional framing of their past criminal offenses more often. These findings highlight a cultural mismatch between minority parole candidates and the expectations of the US justice system, potentially perpetuating racial disparities. Addressing this misalignment may offer a new pathway for reducing systemic bias in discretionary parole decisions.},
    number = {7},
    urldate = {2026-07-01},
    journal = {PNAS Nexus},
    author = {Telidevara, Amrita Kaushik and Boyd, Ryan L and Park, BoKyung},
    month = jul,
    year = {2026},
    pages = {pgag193},
}