Modelpublic

crash-information-extraction/llama3.1-8B-Crash-Location-Casualty

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

Model Card

Llama3.1-8B-Crash-Location-Casualty

Fine-tuned Llama 3.1 for extracting crash location and casualties from Persian social media.

📄 Paper: Extracting traffic crash information from social media: an LLM-based approach – Transportation Letters (2026)

🎯 What it does

Extracts two key types of information from Persian social media texts:

—Location (streets, highways, areas)
—Casualty Count (fatalities & injuries)

⚙️ Fine-tuning

—Method: QLoRA
—Base Model: unsloth/meta-llama-3.1-8b-bnb-4bit
—Data: Proprietary Persian social media crash dataset (Damavand County, Iran)

📊 Performance

Task	Metric	Score
Fatality & Injury Extraction	Accuracy	97%
Location Extraction	BLEU	0.697
Location Extraction	METEOR	0.864

Social media captured 64.1% of crash hotspots compared to official records.

🚀 Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("crash-information-extraction/llama3.1-8B-Crash-Location-Casualty")
tokenizer = AutoTokenizer.from_pretrained("crash-information-extraction/llama3.1-8B-Crash-Location-Casualty")