🗓️ The 30-Day Plan (v2 · baseline-first)
July 3 → August 1. You'll never start from a blank cell — so we don't train like you will.
Why v2? Two facts change everything
- Every IOAI task ships a runnable baseline. Contest day begins with working code that already scores points.
- Gemma 4 writes code for you. You never need to author code — you need to read, assemble, verify, and direct.
The 30 days = this loop on progressively harder baselines. The two skills being trained: diagnosis and Gemma-driving. Python literacy exists only to serve them.
Daily rhythm: every session is LLM-paired under contest rules from Day 1 — one chat = one job, chunked asks ("code only, max 30 lines"), pretend replies are capped, verify by running. ~50% baselines, ~30% LLM-paired reading, ~20% journaling + growing your snippets file. Everything in Kaggle/Colab.
Week 1 · Days 1–7 — Read code before you write code
Day 1 is demystification day: press "Run All" on a finished Titanic notebook, then run the real 2026
Night Watch baseline untouched. You understand nothing yet — fine. Lesson:
the baseline already works; contest day starts from here. Days 2–3: Python by interrogation — paste real
notebook cells into an LLM ("explain line by line"), break the notebook on purpose, read tracebacks bottom-up.
Day 4: numpy + the print(x.shape) ritual. Days 5–6: pandas by asking→reading→running→explaining back.
Checkpoint Day 7: explain every cell of an unseen notebook; make 3 modifications with predicted outcomes.
Week 2 · Days 8–14 — The sklearn improvement loop
Day 8: metric-first reading + first full BIL on Titanic. Day 9: metrics — the confusion matrix is your diagnosis tool. Day 10: validation. Day 11: the 4-model bake-off via ONE chunked prompt. Day 12: feature engineering as loop iterations (change → measure → keep/revert). Day 13: embeddings as a black box + cosine similarity (IOAI's favorite pattern — usable without understanding transformers). Checkpoint Day 14: Antique — full BIL: baseline → diagnose → bake-off → self-training. Pass = beat the baseline and say why.
Week 3 · Days 15–21 — PyTorch = surgery, not authorship
Day 15: tensors + shapes. Day 16: the sacred training loop, assembled from 3 chunked asks (data / model / loop), then explained back. Day 17: surgery drills — head swap, backbone freeze (+ param-count check), loss change, class weights: four moves that solve half of IOAI history. Day 18: the transfer-learning recipe. Day 19: training hygiene + the wiring tests (overfit-10-samples, tiny-slice-first). Day 20: embeddings + Hungarian matching. Checkpoint Day 21: Speech Detector in under 2 hours, contest rules, zero hand-authored code.
Week 4 · Days 22–28 — THE REAL 2026 TASKS + dress rehearsal
Hard rule from here: only whitelisted docs + one LLM chat with self-imposed 2000-token discipline — every line of code comes from the baseline or from Gemma. Days 22–23: Night Watch (watch forgetting happen, then head expansion + replay). Day 24: Robot Delivery (CNN + rare actions + mask). Day 25: John Wilkins (precompute + info gain). Day 26: full 6-hour mock on 3 unseen 2025 tasks — baselines submitted in the first 30 min each? Day 27: post-mortem, grow snippets + prompt bank. Day 28: second pass on your weakest 2026 task.
Days 29–30 — Package & rest
All three 2026 submissions in exact required formats. Reread the rules and the Gemma playbook. Sleep. No new theory.
If you fall behind, never cut: the checkpoints (Days 14, 21), surgery drills (17–18), the three 2026 tasks, the mock (26).