neurontorch/nasa_incident_classifier

NASA Aviation Incident Classifier

This model is a fine-tuned version of microsoft/deberta-v3-large on the NASA Aviation Incident Reports dataset.

It achieves the following results on the evaluation set:

• Loss: 0.3248
• Accuracy: 95%
• F1 Score: 0.9504

Model Description

A state-of-the-art text classification model designed to categorize aviation incident reports into five distinct categories. Built on DeBERTa-v3-large (435M parameters), this model leverages disentangled attention mechanisms for superior natural language understanding of technical aviation terminology.

Classification Categories

Label	Description	Examples
mechanical_failure	Equipment/system malfunctions	Engine failures, hydraulic leaks, electrical issues
human_error	Crew mistakes or procedural errors	Wrong settings, missed checklists, miscommunication
weather_related	Weather-induced incidents	Turbulence, icing, windshear, storms
bird_strike	Wildlife collisions	Bird ingestion, flock encounters
ground_incident	Ground operations issues	Runway incursions, vehicle collisions, ramp incidents

Intended Uses & Limitations

Intended Uses

• Automated classification of aviation safety reports (ASRS-style)
• Safety data analysis and trend identification
• Training data augmentation for aviation NLP research
• Educational demonstrations of transformer-based text classification

Limitations

• Trained on synthetic data modeled after NASA ASRS reports
• Best suited for English-language incident narratives
• May struggle with highly ambiguous multi-cause incidents
• Not validated for operational safety-critical decisions

Training and Evaluation Data

Dataset: neurontorch/nasa_aviation_incident_reports

Split	Samples
Train	400
Eval	100
Total	500

Each category contains 100 balanced samples with professional aviation terminology and root-cause-based labeling.

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: AdamW (fused) with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 20 (early stopping at epoch 10)
• mixed_precision_training: Native AMP (FP16)

Training Results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
1.626	1.0	25	1.5861	0.28	0.1813
1.2794	2.0	50	1.1491	0.57	0.4570
0.9383	3.0	75	0.8447	0.71	0.7000
0.2942	4.0	100	0.3566	0.90	0.9007
0.1401	5.0	125	0.3002	0.94	0.9386
0.1607	6.0	150	0.3129	0.94	0.9402
0.0038	7.0	175	0.3071	0.95	0.9504
0.0022	8.0	200	0.3248	0.95	0.9504
0.0015	9.0	225	0.3409	0.95	0.9504
0.0012	10.0	250	0.3545	0.95	0.9504

Usage

from transformers import pipeline

classifier = pipeline("text-classification", model="neurontorch/nasa_incident_classifier")

# Example inference
text = "Engine oil pressure warning illuminated during cruise at FL350. Crew followed QRH procedures and diverted."
result = classifier(text)
print(result)  # [{'label': 'mechanical_failure', 'score': 0.98}]