3rd Place at AMD AI Sprint Hackathon - IIT Bombay

Our team secured 3rd place at the AMD AI Sprint Hackathon held at IIT Bombay, under Track 1: AMD AI Premier League (AAIPL). The challenge involved building two intelligent agents that battled against each other.

Key Agents

Q-agent: Generates challenging, domain-specific multiple-choice questions.
A-agent: Attempts to accurately answer questions from competing teams.

Main Approach & USPs

Data Sourcing: Sourced high-quality data from domain-relevant books, synthetic datasets, and the OEIS database.
Optimization: Applied research-driven training strategies for model optimization.
Hardware Efficiency: Focused on efficient model finetuning and evaluation within a limited AMD GPU of 192 Gb size for a 24-hour window.

Technical Details

Data Collection: Used structured resources like textbooks, synthetic data generation, and OEIS for number patterns and series.
Finetuning:
- Q-agent: Qwen-3-4B with 50 epochs of supervised finetuning (SFT). Generation speed within 10 seconds per question.
- A-agent: LLaMA-3.2-3B with 50 epochs of SFT. Generation speed within 6-7 seconds per answer.
Prompt Tuning: Incorporated multi-pass retries to enhance both question quality and answer accuracy.