SANDIGAN (Scalable Adaptive Network for Domain-Integrated Governance and Advisory Navigation): Development and Evaluation of a Domain-Fine-Tuned, Internet-free Language Model for Philippine DRRM Advisory Generation

Document Types

Paper Presentation

Research Theme (for Paper Presentation and Poster Presentation submissions only)

Computer and Software Technology, and Robotics (CSR)

School Name

QUEZON NATIONAL HIGH SCHOOL

Track or Strand

Science, Technology, Engineering, and Mathematics (STEM)

Research Advisor (Last Name, First Name, Middle Initial)

TAMBILOC, CARMELA ANA, R.

Start Date

23-6-2026 1:30 PM

End Date

23-6-2026 3:00 PM

Zoom Link/ Room Assignment

DLSU Manila Campus (In-person) - Don Enrique T. Yuchengco Hall - Y503

Abstract/Executive Summary

Cloud-dependent AI tools fail Philippine disaster risk reduction and management (DRRM) operations at the moment they are needed most: active disaster deployments routinely involve power failure and degraded connectivity, rendering internet-dependent advisory systems inaccessible precisely when advisory support is critical. This study addresses that structural incompatibility—termed the connectivity dependency paradox—through three engineering contributions: (1) a reproducible parameter-efficient fine-tuning pipeline that domain-adapts a compact language model (GPT-2, 124M parameters) to Philippine DRRM institutional knowledge in under one GPU-hour on freely available cloud hardware; (2) post-training Q4_K_M quantization that reduces the model from approximately 248 MB to 80 MB (a 68% reduction), enabling on-device mobile inference with no network dependency; and (3) a Flutter-based mobile application implementing a dual-use deployment architecture that serves both PDRRMO institutional workflows and community-facing advisory interfaces from a single quantized model artifact. The fine-tuning pipeline is trained on a structured, barangay-resolution corpus assembled exclusively from authoritative Philippine government sources such as PAGASA, PHIVOLCS, NDRRMC, and the Quezon Provincial DRRMO while covering nine municipalities at the barangay-level hazard resolution. Performance is evaluated against a retrieval-augmented generation (RAG) baseline and a scale ablation using BERTScore F1, protocol compliance rate, a bifurcated hallucination coding scheme, and mean inference latency. Expert practitioner evaluation sessions with six Quezon PDRRMO personnel confirmed the connectivity dependency paradox as an operational reality and independently identified protocol compliance and geographic localization as the critical advisory quality dimensions. Training loss decreased monotonically from 89.24 to 45.35 across three epochs, confirming stable domain adaptation without divergence.

Keywords

domain fine-tuning; low-rank adaptation; on-device inference; retrieval-augmented generation; disaster risk reduction

Statement of Originality

yes

This document is currently not available here.

Share

COinS
 
Jun 23rd, 1:30 PM Jun 23rd, 3:00 PM

SANDIGAN (Scalable Adaptive Network for Domain-Integrated Governance and Advisory Navigation): Development and Evaluation of a Domain-Fine-Tuned, Internet-free Language Model for Philippine DRRM Advisory Generation

Cloud-dependent AI tools fail Philippine disaster risk reduction and management (DRRM) operations at the moment they are needed most: active disaster deployments routinely involve power failure and degraded connectivity, rendering internet-dependent advisory systems inaccessible precisely when advisory support is critical. This study addresses that structural incompatibility—termed the connectivity dependency paradox—through three engineering contributions: (1) a reproducible parameter-efficient fine-tuning pipeline that domain-adapts a compact language model (GPT-2, 124M parameters) to Philippine DRRM institutional knowledge in under one GPU-hour on freely available cloud hardware; (2) post-training Q4_K_M quantization that reduces the model from approximately 248 MB to 80 MB (a 68% reduction), enabling on-device mobile inference with no network dependency; and (3) a Flutter-based mobile application implementing a dual-use deployment architecture that serves both PDRRMO institutional workflows and community-facing advisory interfaces from a single quantized model artifact. The fine-tuning pipeline is trained on a structured, barangay-resolution corpus assembled exclusively from authoritative Philippine government sources such as PAGASA, PHIVOLCS, NDRRMC, and the Quezon Provincial DRRMO while covering nine municipalities at the barangay-level hazard resolution. Performance is evaluated against a retrieval-augmented generation (RAG) baseline and a scale ablation using BERTScore F1, protocol compliance rate, a bifurcated hallucination coding scheme, and mean inference latency. Expert practitioner evaluation sessions with six Quezon PDRRMO personnel confirmed the connectivity dependency paradox as an operational reality and independently identified protocol compliance and geographic localization as the critical advisory quality dimensions. Training loss decreased monotonically from 89.24 to 45.35 across three epochs, confirming stable domain adaptation without divergence.

https://animorepository.dlsu.edu.ph/conf_shsrescon/2026/BoA_CSR/13