So our last training pass on Qwen3-Omni-30B-A3B-Instruct — fine-tuning it on Barbados newspapers — ran out of time halfway through. Twelve hours of GPU budget, planned for 1624 training steps, finished at step 1000. Loss was still falling. The full checkpoint, optimiser state, scheduler — everything you need to keep going — was sitting safely in a private cloud bucket. So we restarted it. Then the interesting bit happened.
The headline finding is the +5 percentage points on the 60-probe text knowledge eval (reproducible to 0pp on the same stack). The bigger finding is the radio audio number, which roughly doubled on completion. Both point to the same conclusion: completing the training schedule to step 1624 mattered.
Context: Future Caribbean and Pulse
This work is part of Future Caribbean, a regional initiative that runs a global agentic AI buildathon over a 21-day sprint each summer. The framing is "build deployable AI systems that strengthen how economies coordinate" — ten tracks, $70K+ in prizes, and a final showcase at the NYSE. We're one of the selected teams.
The programme advertised NVIDIA H200-class compute for selected teams, and we planned this training work on that basis. In practice, compute was not made available to our team during the build window. Rather than drop the model work, I rented the H200 capacity personally through Hugging Face Jobs. That is a decision worth flagging: access to compute materially shapes what small teams can attempt in a three-week buildathon, and going it alone means we could only train the model once, on the budget I could afford.
