dpalate/disaster-tweet-priority-qwen-lora

Disaster Tweet Priority Classifier (Qwen2.5 LoRA)

LoRA adapter for Qwen2.5-0.5B-Instruct, fine-tuned to classify disaster-related tweets into 5 priority levels for emergency triage.

Priority scale

• 1 - Critical: injured/dead, missing persons
• 2 - High: urgent needs, evacuations
• 3 - Medium: infrastructure damage, situational info
• 4 - Low: rescue coordination, advisories
• 5 - Very Low: sympathy, non-humanitarian

Training data

HumAID disaster tweet corpus, remapped from 10 original classes to 5 priority levels. Training set: ~77k tweets across multiple disaster events.

Training config

• LoRA rank 8, alpha 16, dropout 0.05
• Target modules: q/k/v/o_proj, gate/up/down_proj
• 2 epochs, batch size 32, lr 2e-4 cosine
• assistant_only_loss=True (loss only on JSON output)

Output format

The model outputs JSON:

{"priority": 2, "label_name": "High"}

Project

Course project for CS668 - real-time disaster tweet triage with LLMs.