Despite being a high school student, my extensive experience with artificial intelligence drives my commitment to helping technical innovations reach a global audience. This paper presents my approach to directing the LLM track at GIBC 2026 — from designing the 50M parameter challenge to evaluating participant submissions and fostering an environment where novel architectures can emerge.
1. Introduction
I oversee the LLM track at GIBC 2026, a 10-week hackathon where student teams push a 50M parameter transformer to its limits. My work spans architecture design, evaluation pipeline curation, training data preparation, and real-time competitor mentoring.
I am passionate about fostering an innovative environment where ideas can truly flourish. The LLM track is designed to give teams maximum flexibility — architecture, data, training strategy — entirely their call, as long as it stays within the 50M parameter constraint.
2. Model Specification
| Hyperparameter | Value |
|---|---|
| Parameters | 50,000,000 |
| Architecture | Decoder-only Transformer |
| Layers | 8 |
| Attention Heads | 12 |
| Embedding Dim | 768 |
| Training Tokens | 2.4B |
| Optimizer | AdamW (lr 3e-4) |
| Hardware | 2× A100 80GB |
2.1 Architecture Overview
The baseline uses a standard decoder-only layout with pre-norm, rotary positional embeddings, and SwiGLU activations. Teams are evaluated on perplexity, downstream task accuracy, and submission quality.
2.2 Evaluation Criteria
Submissions are scored across three axes: perplexity on a held-out validation set, accuracy on downstream classification tasks, and novelty of the architectural choices made within the parameter budget.
3. Capabilities
4. Key Contributions
Competition Infrastructure. Designed the end-to-end evaluation harness for the LLM track, including baseline model, validation sets, and submission pipeline.
Participant Mentorship. Guide teams through model implementation, hyperparameter tuning, and debugging during the competition.
Community Building. Foster an environment where high school students can engage with cutting-edge AI research in a hands-on, competitive format.
5. Contact & Links
Email: gibc.official.team@gmail.com
Role: Director of LLM — GIBC 2026
Track page: topic-llm.html