🏏 AMD AI Premier League (AAIPL) — Adversarial Agent Orchestration

High-Performance LLM Optimization & Reinforcement Learning on AMD Instinct™ MI300X

---

🧠 1. TEAM ALPHA1

The AMD AI Premier League (AAIPL) is a specialized competitive track hosted at the IIT Delhi Yardi School of AI during the AMD AI Reinforcement Learning Hackathon in February 2026. This implementation focuses on the development of a dual-agent ecosystem: a Q-Agent for adversarial question generation and an A-Agent for deductive reasoning, both specifically optimized for the AMD Instinct™ MI300X ecosystem.

In this "Cricket-style" tournament, models compete in head-to-head innings where performance is measured by an agent's ability to stump opponents with complex puzzles while accurately decoding adversarial inputs under strict latency constraints.

---

🤖 2. The Task: Dual-Agent Architecture

The core objective is to build and optimize two specialized agents designed to interact within an adversarial "Pitcher-Batter" loop:

2.1 The Question Agent (Q-Agent)

• Objective: Generate $N$ puzzle-based questions based on provided topics such as Syllogisms, Seating Arrangements, Blood Relations, and Alphanumeric Series.
• Implementation: Developed in agents/question_model.py and invoked by agents/question_agent.py.
• Format Requirement: Must output questions strictly in the schema specified in sample_question.json.

2.2 The Answer Agent (A-Agent)

• Objective: Solve adversarial questions posed by the opponent's Q-Agent.
• Implementation: Developed in agents/answer_model.py and invoked by agents/answer_agent.py.
• Format Requirement: Must output answers strictly in the schema specified in sample_answer.json.

---

⚙️ 3. Operational Instructions

1. Workstation Access: Sign in to dev.amd-ai-academy.com using the assigned Team ID and Password.
2. Model Protocols: Use only authorized models provided in /root/.cache/huggingface/hub.
3. Modification Rule: Hub models are read-only; they must be copied into the AAIPL/hf_models folder for editing. Modifying the original folder results in immediate disqualification.
4. Synchronization: Coordinate with team members to ensure simultaneous notebook edits do not overwrite work.
5. Submission: Push all code (excluding hf_models) to GitHub using the git.sh script before the deadline.

---

🏟️ 4. Tournament Overview & Scoring

All matches are 1v1 knockout format where two teams switch sides in a two-inning structure.

4.1 Innings Structure

• 1st Inning: Team-A (Q-Agent) pitches $N$ questions $\rightarrow$ Team-B (A-Agent) solves.
• 2nd Inning: Team-B (Q-Agent) pitches $N$ questions $\rightarrow$ Team-A (A-Agent) solves.

4.2 Scoring Criteria

Performance is measured by competitive accuracy using the following formulas:

$$\text{A-agent Score} = \dfrac{\text{Count of questions correctly answered in expected format}}{N} \times 100$$

$$\text{Q-agent Score} = \dfrac{\text{Count of questions incorrectly answered by opponent}}{N} \times 100$$

Format Integrity Rule: Teams must maintain a minimum 50% format-correctness rate to avoid automatic disqualification. In case of a TIE, closed benchmark questions are used to evaluate A-Agents.

---

📋 5. Guidelines

Only responses from the Q-Agent and A-Agent that strictly follow the JSON formats below will be considered for evaluation.

5.1 Q-Agent Format (Output Schema)

{
    "topic": "<Topic of the Question>",
    "question": "<full question text>",
    "choices": [
        "A) <choice A text>",
        "B) <choice B text>",
        "C) <choice C text>",
        "D) <choice D text>"
    ],
    "answer": "<correct choice letter only>",
    "explanation": "brief explanation within 100 words"
}

5.2 A-Agent Format (Output Schema)

{
    "answer": "<correct choice letter only>",
    "reasoning": "brief reasoning within 100 words"
}

---

🛠️ 6. Submission & Persistence

• Workspace: All work must be contained within the AAIPL folder.
• Execution: Agents are invoked via python -m agents.question_agent and python -m agents.answer_agent.
• Persistence: Inference results must be saved specifically to outputs/questions.json and outputs/answers.json.
• Checkpoints: Ensure model checkpoints (e.g., .safetensors, .pt) load correctly during automated evaluation.

---

⚠️ 7. Strict Restrictions

Failure to comply results in immediate disqualification:

• No RAG: Retrieval Augmented Generation is strictly prohibited.
• Adversarial Integrity: Strategies designed to force opponent "hallucinations" are disallowed.
• Language: Strictly English only for both agents.
• Latency SLAs: * Question Generation: Under 13 seconds per question.
  • Answer Generation: Under 9 seconds per answer.

---

📂 8. Directory Overview

.
├── agents/
│   ├── question_model.py      # Core Q-agent logic
│   ├── question_agent.py      # Inference driver for Q
│   ├── answer_model.py        # Core A-agent logic
│   └── answer_agent.py        # Inference driver for A
├── assets/
│   ├── topics.json            # Target topics for generation
│   ├── sample_question.json   # Q-format specification
│   └── sample_answer.json     # A-format specification
├── qgen.yaml / agen.yaml      # Generation parameters
└── README.md                  # Project Dashboard

---

🚀 9. Getting Started

To verify your system on the AMD Instinct™ MI300X workstation:

Generate Questions

python -m agents.question_agent --output_file "outputs/questions.json" --num_questions 20 --verbose

Evaluate Answers

python -m agents.answer_agent --input_file "outputs/filtered_questions.json" --output_file "outputs/answers.json" --verbose

---

🔬 Acknowledgments

Developed for the AMD AI Reinforcement Learning Hackathon (Feb 2026) at IIT Delhi. Special thanks to the AMD Engineering Team for MI300X compute access and Unsloth for performance optimization support.