Competitions / 2020 – 2026
Rankings, with the receipts.
Only results I can point to a leaderboard or submission for. The older entries are historical; the 2026 Kaggle entries are works in progress, not medal claims.
01 / Results
Selected competitions.
The useful output is not the certificate. It is the method, the failure mode and the thing I would do differently next time.
| Year / event | Task | Approach | Result |
|---|---|---|---|
| 2026Kaggle · Name That LLM | 18-way attribution of passages to model variants. | TF-IDF baseline → feature, SVM and hierarchical experiments → Qwen3-1.7B LoRA with prompt-level splits and seed checks. | 190k passages; +0.026 balanced accuracy from data cleaning. Research note ↗ |
| 2026Kaggle · Physical-Engineering's Last Exam | Build an autonomous agent harness for 85 long-horizon structural, civil, electrical and mechanical engineering tasks. | Minimal WebSocket harness; hard per-task time budget; input digest with a document-conflict scan; BM25 clause retrieval; IFC clash and connectivity checker. | 5-task dev set at 0.38 mean over two rounds, ¥0.49 per task. Writeup in progress; no rank claimed. Research note ↗ |
| 2026Kaggle · Gemma 4 Developer Agent | Post-train Gemma 4 into an autonomous software-engineering agent that patches real repositories. | Static submission bundle; thinking-budget, per-task time-cap and tool ablations; edit and code-graph tools; forked public ensemble stack. | 0.18 public with the own harness; 0.664 reproducing the public stack. No prize rank. |
| 2026Kaggle · Gemma 4 Prompt Optimization | Prompt optimization against a fixed evaluation harness. | Multiple notebook iterations, topic scraping, a public Dockerfile and repeated submission cycles. | In progress. No rank or medal claimed. |
| 2020中国大学生保险数字科技挑战赛 | Dialogue intent recognition from anonymized character and word IDs. | Word2vec 128d, Transformer + BiLSTM, focal/F1 hybrid loss for class imbalance. | Rank 14 overall · 4th in East China · 0.29306 macro F1. |
| 2020天池 · 饿了么骑士行为预估 | Predict a courier's next action during COVID-19. | LightGBM with wave, order and distance features, extended from a public baseline. | 45 / 102 · 0.751 online score. |
| 2020Tencent Advertising Algorithm Competition | Infer audience attributes from sparse ad-interaction behaviour. | Anonymised interaction features; round-1 submission. Full method not preserved. | Round-1 rank 51 / 1,000+ teams. |
NoteHistorical results are shown as recorded at the time. Current Kaggle entries do not claim a leaderboard position.
02 / Takeaways
What the competitions changed.
Three habits that survived the leaderboard and made the game work better.
Align the metric
Macro F1, balanced accuracy and online score reward different behaviour. Pick the loss and split that match the metric, not the leaderboard folklore.
Build the split first
Prompt-level splits and seed checks catch more leakage than another model. A clean negative result is information.
Keep the failures
The feature that did not replicate and the loss that did not help are part of the result. They are also the most useful interview material.