000
GIBC V2
Global Innovation Build Challenge
GIBC V2 · TRACK 01

LLM (50M PARAMS)

Build a language model from scratch — 50 million parameters max. Architecture, data, and training — entirely your call. Competition duration. Go.

PRIZES TBA OPEN TEAMS COMPETITION
50M
PARAMS

PRIZES

TBA

LLM

TRACK OVERVIEW

01

PRE-TRAIN

Curate your dataset, design your architecture, and pre-train an LLM from scratch. Focus on data quality and training efficiency.

02

EVALUATE

Models are evaluated on perplexity, reasoning benchmarks, and efficiency metrics. Novel approaches earn bonus points.

03

PRESENT

Present your approach in a 5-minute lightning talk. Explain your architecture, training strategy, and key learnings.

CHALLENGE DETAILS

THE PROBLEM

Language models are transforming every industry, but training them remains inaccessible to most. GIBC V2 challenges you to build a competitive model — pushing the boundaries of what small-scale LLMs can achieve.

CONSTRAINTS

  • Maximum 50M parameters
  • Open training window
  • Open-source frameworks only
  • No team size limit

NO PROVIDED RESOURCES

Bring your own infrastructure. No GPUs, no datasets, no baseline models provided. Your setup, your stack, your rules — build however you want.

SKILLS NEEDED

Python PyTorch Transformers NLP Data Curation CUDA HuggingFace W&B

HACKATHON TIMELINE

DAY 1 · 9:00 AM

Kickoff & Briefing

Track intro, rules, Q&A. Hacking begins at noon.

DAY 1 · 12:00 PM

Hacking Begins

10-week clock starts. Build your stack, train your model, break things.

DAY 1–2

Mentor Office Hours

Scheduled check-ins with ML engineers for guidance on training, debugging, and optimization.

DAY 2 · 12:00 PM

Training Deadline

Submit models for evaluation. Leaderboard updates in real-time.

DAY 2 · 4:00 PM

Presentations & Awards

5-min lightning talks. Winners announced at closing ceremony.

EVALUATION CRITERIA

WEIGHT · 40%

Perplexity & Accuracy

Language modeling quality on held-out validation sets. Lower perplexity and higher benchmark accuracy score higher.

WEIGHT · 25%

Reasoning Performance

Performance on logic, math, and commonsense reasoning tasks. Evaluated through standardized benchmarks.

WEIGHT · 20%

Training Efficiency

Compute efficiency, training speed, and memory footprint. Efficient architectures and training strategies rewarded.

WEIGHT · 15%

Innovation

Novel architecture decisions, creative data strategies, or unique training approaches. Bonus for originality.

PRIZES & REWARDS

COMING
SOON

Prize details are still being finalised. Check back soon — we're putting together something worth building for.

READY TO BUILD?

ENTER THE
LLM TRACK

Bring your own stack, build your own model. Competition duration. No limits. Registration closes July 10, 2026.

LLM · 50M PARAMS · PRIZES TBA · 10 WEEKS · JULY 2026 · LLM · 50M PARAMS · PRIZES TBA · 10 WEEKS · JULY 2026 ·