arielcerdap/modernbert-disfluency-expB-mixed

ModernBERT Disfluency Detection — Exp B (Mixed 80/20)

Fine-tuned from answerdotai/ModernBERT-base on mixed data (80% synthetic / 20% real) from FluencyBank Timestamped.

Dataset

• Config: mixed_8020 de arielcerdap/disfluency-fluencybank
• Train: 13713 segmentos (80% sint / 20% real)
• Val/Test: idénticos a Exp A para comparación directa

Labels

O · FP (filled pause) · RP (repetition) · RV (revision) · PW (partial word)

Test Results

Label	P	R	F1	Support
Label	P	R	F1	Support
---	---	---	---	---
O	0.9710	0.9590	0.9650	3704
FP	0.9832	1.0000	0.9915	176
RP	0.6414	0.7299	0.6828	174
RV	0.1748	0.2907	0.2183	86
PW	0.9744	0.8155	0.8879	233

Macro F1 (4 disfluencias): 0.6951
Binary F1: 0.8006

Hyperparameters

• learning_rate: 5e-05
• epochs: 15
• warmup_steps: 963
• weight_decay: 0.1
• focal_loss_gamma: 3.0 (adaptive)
• class_weights: O=1.0, FP=3.0, RP=6.0, RV=20.0, PW=5.0