skip to content
jasondank.com/projects/sailing-rules-assistant

Projects / 2024 · 4 of 13

AI Model Fine-Tuning

Sailing Rules Assistant

train.json{ "instruction": "Who keeps clear?", "context": "Port vs starboard", "output": "Port. Rule 10"},{ "instruction": "Who gets mark-room?", "context": "Overlapped at zone", "output": "Inside boat. Rule 18"},{ "instruction": "Same tack, overlap?", "context": "Windward/leeward", "output": "Windward. Rule 11"},{ "instruction": "I touched a mark?", "context": "Rounding the mark", "output": "One turn. Rule 31"},{ "instruction": "Who keeps clear?", "context": "Port vs starboard", "output": "Port. Rule 10"},{ "instruction": "Who gets mark-room?", "context": "Overlapped at zone", "output": "Inside boat. Rule 18"},{ "instruction": "Same tack, overlap?", "context": "Windward/leeward", "output": "Windward. Rule 11"},{ "instruction": "I touched a mark?", "context": "Rounding the mark", "output": "One turn. Rule 31"},FINE-TUNING FLAN-T5forward →← update weightsTRAINING LOSSepoch 1/3epoch 2/3epoch 3/3SAILING RULES ASSISTANTRacing Rules of Sailing 2025–2028model readyI’m on port, they’re onstarboard and converging.Who keeps clear?Rule 10: on opposite tacks, aport-tack boat shall keep clearof a starboard-tack boat.
Instruction records flow forward through the network, backprop updates the weights layer by layer, the loss comes down, and the assistant answers a rules question

What it does

Fine-tuning a Flan-T5 model for answering questions about sailing rules and interpreting racing scenarios. The model is trained on JSON data containing instructions, optional context, and expected outputs to create an assistant that can answer questions about sailing rules.

  • Designed and implemented a Flan-T5 fine-tuning solution (Python-only codebase) to answer sailing rules questions and interpret racing scenarios.
  • Trained on structured JSON of instructions, optional context, and expected outputs; wrote the training and evaluation pipelines.

What I'm proudest of

It brought together the two things I spend the most time on, racing and machine learning, and I wrote the training and evaluation pipelines end to end.

Screenshots

The Sailing Rules Assistant (2025–2028 rules): a question, optional context and the answer panel, with example questions below
The Sailing Rules Assistant (2025–2028 rules): a question, optional context and the answer panel, with example questions below