Discord

Checklist IOAI PH 2026 Team Selection · Task 1

The "Bayanihan" Crisis Concept Extractor

Extract urgency, location and rescue/medical needs from 10,000 Taglish disaster SMS messages and 50 noisy radio audio clips using ASR and free-tier LLM APIs.

  • Multimodal
  • Information extraction (text and speech) with LLMs

The task

During severe weather, rescue organisations receive large numbers of unstructured distress messages. The dataset contains 10,000 synthetic records: Taglish SMS texts with heavy abbreviations, slang and noise (spam, wrong numbers), and 50 shortwave radio intercepts (.mp3) with dispatcher 10-codes over static.

Contestants write a pipeline that transcribes the audio programmatically (for example with Whisper), queries a free-tier LLM API (for example OpenRouter or Groq) with few-shot prompts, and outputs four labels per record: urgency_level (Critical, High, Moderate, Non-Emergency), location (string or 'Unknown'), needs_rescue (Yes/No) and needs_medical (Yes/No).

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
test.csv (id, raw_text, audio_filename), 50 audio files, benchmark_submission.csv.
You submit
A 'melted' CSV of exactly 40,000 rows: RowId (e.g. 1_urgency_level) and Prediction.
Scoring
Macro F1-score averaged across the four label types.
Rules
  • All transcription and extraction must run programmatically in the Kaggle notebook; no manual transcriptions, hard-coded answers or web GUIs.
  • Data only from the official competition input directory.
  • Individual work; API keys must be loaded via Kaggle Secrets, never hard-coded.
  • Notebooks must be made public or shared with the organisers after the deadline.
Format
IOAI Philippines 2026 Team Selection, Task 1 (private Kaggle community competition).

Details

Year
2026, Online (private Kaggle community competitions)
Round
Team Selection · Task 1
Language
English
License
MIT (repository LICENSE), as stated by the source