sludge-process-roberta

Fine-tuned RoBERTa-base classifier for detecting process sludge (barriers to action — excessive procedural friction) in consumer financial complaint narratives.

Developed and validated in:

Chesterfield, A., Gillespie, A., Goddard, A. and Krpan, D. (2026). Feeling the Friction: Developing and validating text classifiers for sludge in consumer complaints. (manuscript in preparation)

Code and data: GitHub | Companion model: sludge-informational-roberta

What this model detects

Process sludge = barriers to action: excessive procedural friction such as repeated documentation requests, unnecessary steps, or being passed between departments without resolution. Binary classification: 1 = process sludge present, 0 = absent.

Training details

Parameter	Value
Base model	roberta-base
Training samples	825 (85% stratified split of n=971)
Test samples	146 (15% held-out, evaluated once)
Epochs	4
Learning rate	2e-5
Batch size	8
Warmup ratio	0.1
Class weighting	Balanced
Random seed	42

Performance (held-out test set, n=146)

Metric	Score
MCC	0.796
F1	0.863
Precision	0.800
Recall	0.936
Accuracy	0.904

How to use

from transformers import pipeline

classifier = pipeline("text-classification", model="AlexChesterfield/sludge-process-roberta")
complaint = "I called five times and each time was transferred to a different department."
result = classifier(complaint)
# LABEL_1 = sludge present, LABEL_0 = absent

Citation

@article{chesterfield2026sludge,
  title={Feeling the Friction: Developing and validating text classifiers for sludge in consumer complaints},
  author={Chesterfield, Alexandra and Gillespie, Alex and Goddard, Alex and Krpan, Dario},
  year={2026}
}