---
language: en
tags:
- disfluency-detection
- token-classification
- modernbert
- speech-pathology
datasets:
- arielcerdap/disfluency-fluencybank
---

# ModernBERT-large Disfluency Detection — Exp C (Real Only)

Fine-tuned from [answerdotai/ModernBERT-large](https://huggingface.co/answerdotai/ModernBERT-large) on **real-only**
FluencyBank Timestamped data. Identical setup to Exp A (base) to isolate architecture effect.

## Dataset
- Config: `real_only` de `arielcerdap/disfluency-fluencybank`
- Train: 2737 segmentos (100% reales)
- Val/Test: idénticos a Exp A y B para comparación directa

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

## Test Results vs Exp A (base)
| Label | Exp A (base) | Exp C (large) |
|---|---|---|
| FP | 0.9944 | 0.9944 |
| RP | 0.8022 | 0.8964 |
| RV | 0.3145 | 0.4974 |
| PW | 0.8879 | 0.9451 |
| Macro (dis) | 0.7497 | 0.8333 |
| Binary F1 | 0.8902 | 0.9250 |

## Per-class Detail
| Label | P | R | F1 | Support |
|---|---|---|---|---|
| O | 0.9892 | 0.9846 | 0.9869 | 3704 |
| FP | 0.9888 | 1.0000 | 0.9944 | 176 |
| RP | 0.8743 | 0.9195 | 0.8964 | 174 |
| RV | 0.4563 | 0.5465 | 0.4974 | 86 |
| PW | 0.9685 | 0.9227 | 0.9451 | 233 |


## Hyperparameters
- learning_rate: 5e-05
- batch_size effective: 32 (8 × 4 grad_accum)
- epochs: 15
- warmup_steps: 191
- weight_decay: 0.1
- classifier_dropout: 0.3
- focal_loss_gamma: 3.0 (adaptive)
- class_weights: O=1.0, FP=3.0, RP=6.0, RV=20.0, PW=5.0