gregoire-marie/xlm-roberta-atc

ATC Severity Classifier

Transformer classifier for ATC transcript severity detection with three labels: NORMAL, URGENCY, and DISTRESS.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("gregoire-marie/xlm-roberta-atc")
model = AutoModelForSequenceClassification.from_pretrained("gregoire-marie/xlm-roberta-atc")

Label Set

• NORMAL
• URGENCY
• DISTRESS

Training Configuration

• Base model: xlm-roberta-base
• Max length: 128
• Epochs: 5.0
• Train batch size: 16
• Train/val/test sizes: 700 / 150 / 150

Latest Test Metrics

• Accuracy: 0.96
• Recall (URGENCY): 1.0
• Recall (DISTRESS): 0.8181818181818182
• False alarm rate (NORMAL -> emergency): 0.022222222222222223

Per-Class Report

• URGENCY precision/recall/f1: 0.8837209302325582 / 1.0 / 0.9382716049382716
• DISTRESS precision/recall/f1: 1.0 / 0.8181818181818182 / 0.9

Files

• Uploaded from /Users/gregoire/src/atc-severity-classifier/outputs/transformer/run_20260327_000927/best_model
• The repository contains a Hugging Face-compatible config.json, tokenizer files, and model weights.