Participant GuideHow to
How-to — recipes for specific tasks
Sourced from whest-starterkit @
96e308c.
How-to — recipes for specific tasks
Task-oriented guides. Each one answers "I want to do X — give me the steps and the common mistakes." Use these alongside the Tutorial for a guided sequence and the Reference for exact APIs.
Writing and iterating
| Doc | When to read |
|---|---|
| write-an-estimator.md | Implementing your custom estimator from scratch: minimal structure, contract checklist, common first failure. |
| inspect-mlp-structure.md | Traversing the MLP object: fields, weights, shapes. |
| validate-run-package.md | The standard local loop: validate → run → package. Includes a "Useful flags" table covering --seed, --n-samples, --profile, --max-threads, --setup-timeout, and other whest run flags. |
| use-evaluation-datasets.md | Pre-creating an evaluation dataset for fast, reproducible iteration. |
Optimizing
| Doc | When to read |
|---|---|
| algorithm-ideas.md | Survey of estimation strategies: Monte Carlo, mean propagation, covariance, and why one fixed strategy outperforms runtime routing at this budget. Plus open directions (low-rank, layer-adaptive, spectral, importance sampling, higher moments). |
| manage-flop-budget.md | Where your FLOPs go and how to fit a tighter budget. Includes a line-by-line walkthrough of examples/02_mean_propagation.py. |
| performance-tips.md | Concrete patterns: matmul placement, free ops, diagonal vs covariance, env-var knobs. |
Debugging and shipping
| Doc | When to read |
|---|---|
| debugging-checklist.md | Tiered checklist for "estimator runs but something feels wrong": Tier 0 pure-Python loop, Tier 1 sanity, Tier 2 correctness, Tier 3 optimization. |
| pre-submission-checklist.md | Checks to run before you click "submit" on AIcrowd, including the Phase 2 allowed-code rules and the 10-submissions-per-UTC-day cap. |
| ship-weights.md | Pre-compute offline and load weights in setup() via submission_dir; the boundary between permitted data and prohibited code; the 5 s setup budget; multi-file submissions; 50 MiB / 50 file caps; .whestignore; package preview and --dry-run. |
➡️ Where to look next
- Not sure whether a technique is allowed at all? → Allowed Code.
- Need the exact contract / report fields? → Reference.
- Estimator throws an error you don't recognize? → Troubleshooting.
- Working through the stages in order? → Tutorial.