IOAI 2024 · Scientific Round (set)
The Scientific Round of the inaugural IOAI (Burgas, Bulgaria, August 2024) ran in two stages: a one-month at-home stage with three full ML/NLP/CV tasks, and an 8-hour on-site stage that mutated each task into a harder variant under contest time pressure. This set covers all four walkthroughs (3 at-home + 1 on-site twin, with the other on-site siblings folded into each task page).
Round metadata
| Year | 2024 (1st edition) |
|---|---|
| Host | Burgas, Bulgaria · 9–15 August 2024 |
| Round | Scientific (at-home + on-site) |
| Tasks | 3 at-home (ML, NLP, CV) + 3 on-site siblings |
| Duration | At-home: 1 month · On-site: 8 hours |
| Hardware | Single L4 GPU (HuggingFace Spaces / Kaggle equivalent) |
| Allowed | Any open-weights model from HuggingFace, PyTorch, standard scientific Python stack |
| Scoring | Per-task metric scaled to 0–100; total is the sum (max 300 for Scientific) [illustrative weighting] |
Tasks
Feature engineering on matrix-shaped samples
Generate the best feature representation for a frozen downstream model, where each training example is itself a small matrix. On-site sibling flips the task type and the dataset.
Fine-tune a model on a ciphered language
5-way text classifier in an invented language. On-site sibling extends the head to 7 classes with no new learned parameters. (Existing walkthrough — already in the archive.)
Swap zebras and giraffes by editing SDXL-mini weights
Diffusion-model weight surgery: same prompt should now produce the swapped animal. On-site sibling: insert a hydrant alongside the cow without breaking the original prompt.
Cow + hydrant: compose two concepts via weight edits
8-hour on-site variant of the CV at-home task. Tighter time budget forces a different family of edits (textual-inversion vs ROME-style updates).
How to use this set
- Read the parent (this page) once, then pick one task and treat it as a timed mock — at-home tasks are designed to take days, but a focused 4-hour session is enough to reach a solid mid-board score.
- Before peeking at a walkthrough, write down your own baseline plan. Compare which "improvement levers" you missed — that diff is the actual learning signal.
- For the on-site sibling, set an 8-hour total budget and force yourself to ship a submission file every 2 hours so you always have a fallback.
- Hardware-wise, the L4 budget is real: a Colab T4 is close enough for practice. If you can't reproduce in < 1 hour on a free GPU, your architecture is too big.