ModernBERT-base Disfluency Detection — Real Data Baseline
Fine-tuned from answerdotai/ModernBERT-base using only real data from FluencyBank Timestamped (Romana et al., 2024).
Purpose
This model serves as the Experiment A baseline in an ablation study comparing:
- This model: trained on real data only (2,744 train examples)
- Mixed model: trained on 80% synthetic + 20% real (13,713 train examples)
The comparison quantifies the contribution of the synthetic data augmentation pipeline.
Dataset
FluencyBank Timestamped — 3,430 segments from 37 adults who stutter. Split: 80/10/10 train/val/test (random_state=42). No synthetic data used.
Label Priority
FP > PW > RP > RV (corrected from original FP > RP > RV > PW) This allows ~2,048 real PW tokens to be correctly labeled.
Test Results
- Overall Accuracy : 0.9588
- Overall F1 (macro): 0.8312
- FP F1: 0.0000
- RP F1: 0.0000
- RV F1: 0.0000
- PW F1: 0.0000
- Downloads last month
- 7