Upload 2 files

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+wsl-v4.sqlite filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,236 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+language:
+- en
+pretty_name: WSL Championship Tour Heat-Level Dataset (2014–2025)
+size_categories:
+- 100K<n<1M
+task_categories:
+- tabular-regression
+- tabular-classification
+- time-series-forecasting
+- other
+tags:
+- sports
+- surfing
+- wsl
+- world-surf-league
+- competition
+- fantasy
+- wave-pool
+- ocean
+- era5
+- weather
+- analytics
+configs:
+- config_name: default
+  data_files:
+  - split: full
+    path: wsl-v4.sqlite
+extra_gated_prompt: >-
+  This dataset is released for non-commercial research and analytical use under
+  CC BY 4.0. Surf scoring data is scraped from public WSL results pages and is
+  not an official WSL product. ERA5 weather data is from Open-Meteo under CC BY 4.0.
+---
+
+# WSL Championship Tour Heat-Level Dataset (2014–2025)
+
+A complete, structured, hash-chained dataset of every World Surf League Men's and Women's Championship Tour heat from 2014 through 2025 — 230 events, 7,494 heats, 17,710 surfer-heat rows, 103,489 individual waves — joined to break geometry and ERA5 oceanic/atmospheric reanalysis.
+
+Built for fantasy-league optimization, surf analytics, and educational machine-learning work. Every row is identified by a deterministic SHA-256 hash chain, every value is sourced from the public WSL website with documented provenance, and every known source-data limitation is recorded explicitly rather than papered over.
+
+---
+
+## Quick start
+
+```python
+import sqlite3
+import pandas as pd
+
+conn = sqlite3.connect("wsl-v4.sqlite")
+
+# All heats from a recent event with conditions
+heats = pd.read_sql("""
+    SELECT e.name, h.heat_date, h.wave_range, h.wind_condition,
+           era5.wave_height_m, era5.wind_speed_kmh
+    FROM Event e
+    JOIN Heat h ON h.event_id = e.event_id
+    LEFT JOIN ERA5Conditions era5
+      ON era5.break_id = h.break_id
+     AND DATE(era5.observation_hour) = h.heat_date
+     AND CAST(substr(era5.observation_hour, 12, 2) AS INTEGER) = 12  -- noon
+    WHERE e.year = 2025 AND e.tour = 'MCT'
+""", conn)
+```
+
+For a working schema dump, run `sqlite3 wsl-v4.sqlite '.schema'` after download.
+
+---
+
+## Dataset summary
+
+- **Coverage:** 2014–2025, both Men's CT (MCT) and Women's CT (WCT), all events including wave-pool stops (Surf Ranch, Freshwater Pro, Surf Abu Dhabi, Jeep Surf Ranch).
+- **Granularity:** Wave-level. Every wave a surfer caught in every heat is a row, with score, counted/uncounted flag, interference flag, and direction (for wave-pool events).
+- **Format:** Single SQLite file (~18 MB, 9 tables, 7 indexes). The compact size is deliberate — see "ERA5 design choice" below.
+- **Identity:** Every primary key is a deterministic SHA-256 hash truncated to 16 hex characters. The hash chain encodes ancestry: `wave_id = SHA256(heat_id|athlete_id|wave_index)`, `heat_id = SHA256(round_id|heat_number)`, etc. Two independent observers computing IDs from the same source HTML produce the same IDs — the dataset is reproducible by construction.
+- **Conditions:** ERA5 reanalysis from Open-Meteo (Marine + Archive APIs) joined by `(break_id, observation_hour)`. 98.5% of heats join to a full 24-hour ERA5 day at their break.
+- **Audit:** Verdict HEALTHY, 0 Critical, 0 Major findings. 365 / 365 stratified-sample heats verified clean against the live WSL site at publication time.
+
+---
+
+## Tables
+
+| Table | Rows | Purpose |
+|---|---|---|
+| `Event` | 230 | One row per (year, tour, name, location, anomaly_class). MCT and WCT are always separate rows. |
+| `Round` | 1,496 | Three columns: `round_id`, `event_id`, `round_number`. No display name (app-layer concern). |
+| `Heat` | 7,494 | The principal competitive unit. Date, duration, conditions, break, `wsl_heat_id` for source traceability. |
+| `HeatSurfer` | 17,710 | (heat × surfer) pairing with `total_score` and `outcome_type`. |
+| `Wave` | 103,489 | One row per wave. Score, `is_counted` (top-2-counted indicator), `is_zeroed_by_interference`, `direction` (wave-pool only). |
+| `Surfer` | 474 | Surfer profile: nationality, stance, DOB, height/weight (where WSL publishes them). |
+| `Break` | 40 | Geographic + competitive metadata — lat/lon, break type/handedness, swell window, orientation. |
+| `BreakAlias` | 196 | Year-aware mapping of WSL-rendered location strings → canonical `break_id`. |
+| `ERA5Conditions` | 15,264 | Hourly ERA5 reanalysis: wave height/period/direction, wind speed/direction, surface pressure. |
+
+### Hash chain (governing identity rule)
+
+```
+event_id     = SHA256(name | year | tour | location | x)[:16]
+round_id     = SHA256(event_id | round_number)[:16]
+heat_id      = SHA256(round_id | heat_number)[:16]
+heatentry_id = SHA256(heat_id | athlete_id)[:16]
+wave_id      = SHA256(heat_id | athlete_id | wave_index)[:16]
+break_id     = SHA256(name | lat:.6f | lon:.6f)[:16]
+condition_id = SHA256(break_id | observation_hour)[:16]
+athlete_id   = SHA256(name | date_of_birth | nationality)[:16]
+```
+
+`x = 1` for multi-break events (heats span breaks with different competitive character — e.g. 2018 Margaret River split across North Point, Margaret River, Uluwatu after a shark incident; 2020 Maui Pro split between Honolua Bay and Pipeline). Heat-level `break_id` is authoritative for which specific break each heat ran at.
+
+NULL sentinel: any field that is NULL in the source uses the literal string `"NULL"` in hash inputs. Exception: `heat_number` falls back to 1-based positional ordering within its round when the source HTML doesn't provide a number, to avoid hash collisions.
+
+---
+
+## Coverage by year and tour
+
+| Year | MCT events | WCT events | Notes |
+|---|---|---|---|
+| 2014–2019 | 11 | 10 | Standard tour |
+| 2020 | 1 | 1 | COVID-disrupted year. Only the December 2020 Pipe Masters and Maui Pro made it. |
+| 2021 | 7 | 7 | Truncated COVID-recovery season |
+| 2022–2024 | 10–11 | 10–11 | Standard tour |
+| 2025 | 12 | 12 | Current full season (latest) |
+
+**Pre-2014 data is not included in this release.** Earlier development generations of this dataset once contained partial MCT 2008–2013 and WCT 2010–2013 data scraped from WSL's archive, but per-wave detail is incomplete for that era and those generations were superseded before v4. The current production database and the available pre-v4 backup both cover 2014 onward only. v4 is therefore *not* a 2008-onward archive with the early years stripped — it is a uniform 2014–2025 dataset by design, where every event has full wave-level fidelity.
+
+---
+
+## Data sources and provenance
+
+**Competition data** is scraped from public WSL results pages (`worldsurfleague.com/events/...`) using the heat-popup endpoint and round-listing pages. Every heat record is traceable back to a specific WSL URL via `Heat.wsl_heat_id`. Recent re-crawls (April 2026) used direct popup-by-popup parsing with grid-attribution validation; earlier crawl batches were verified end-to-end against the same source.
+
+**ERA5 conditions** come from Open-Meteo's Marine API (`models=era5_ocean`) and Archive API (`models=era5`). Hourly resolution, 0.25–0.5° spatial. Stored as raw values — no aggregation at the storage layer.
+
+**ERA5 design choice — why the dataset is ~18 MB.** A naïve enrichment fetches ERA5 for every hour of every day across each break's full historical span. For 28 ocean breaks over a 12-year window that's ~2.9 million hourly rows, the vast majority of which would never join to a heat (heats happen on a handful of discrete days per year, not continuously). The v3 predecessor of this dataset stored ~129 MB primarily for that reason. v4 instead fetches ERA5 only for the actual heat days each break sees — one 24-hour ERA5 day per unique `(break_id, heat_date)` pair — producing 15,336 rows with **100% ocean-heat join coverage**. Same usability, ~7× smaller file. Wave-pool venues (Surf Ranch, Surf Abu Dhabi) carry zero ERA5 rows by policy: pool conditions are mechanically generated and are not derivable from offshore reanalysis, so populating them would create joinable but physically meaningless values.
+
+**Surfer profile data** (nationality, stance, DOB, height/weight) is scraped from individual athlete profile pages. Nationality follows WSL's competitive designation rather than legal citizenship (e.g. Hawaiian surfers are listed as "Hawaii", not "United States" — WSL treats Hawaii as a separate competitive territory).
+
+**No carry-forward.** Every row of competition data was fetched fresh from WSL's current servers in 2026, parsed by an audited parser, and validated against per-heat consistency rules (e.g. `total_score == sum(top-2 counted scores)` for every surfer). Six events that were originally carried forward from a prior release were purged and re-crawled in this release.
+
+---
+
+## Verification status
+
+- **Audit verdict:** HEALTHY (0 Critical, 0 Major, 14 Minor, 36 Info findings).
+- **Stratified sample verifier:** 365 / 365 heats verified clean. Sample stratifies by year (≥15 ocean heats per year) and per wave-pool event (≥3 heats per event, or all heats for events with fewer than 3 by structure).
+- **Per-heat consistency:** Every `HeatSurfer.total_score` reconciles with `SUM` of the top-2 counted Wave scores within the same (heat, athlete) within ±0.011 tolerance, except where WSL applied an interference penalty.
+- **Hash-chain integrity:** Every `Wave` traces to a `HeatSurfer` traces to a `Heat` traces to a `Round` traces to an `Event`; no orphan rows.
+
+---
+
+## Known limitations (documented)
+
+1. **Heat times not stored.** Granularity is day-only (`heat_date`). WSL does not publish per-heat start times for historical heats. ERA5 joins use day-level matching and let the consumer decide which hour(s) to interpolate.
+
+2. **Penalty 1 (interference halving) not detectable.** WSL applies interference at three tiers: (1) second-best wave halved, (2) second-best wave zeroed, (3) best wave zeroed. Tiers 2 and 3 are captured via `Wave.is_zeroed_by_interference = 1` (cell carries `wave--adjusted` class AND score = 0.0). Tier 1 is a score reduction with no machine-readable HTML signal — the pre-halving score is never published. The post-halving score (which we store) is correct as a score; only the "this was halved" annotation is missing. **No re-crawl will fix this — it's a source-data limitation.**
+
+3. **TBD waves.** WSL renders some wave-grid cells as the literal text "TBD" — slots that exist on the wave grid but never received a numeric score (likely WSL data-entry gaps). v4 stores TBD waves as `Wave` rows with `score = NULL`, `is_counted = 0`. The wave_count column matches WSL's display; aggregations should filter `WHERE score IS NOT NULL` to exclude. There are 12 TBD waves total.
+
+4. **`Heat.wsl_heat_id = -1` sentinel.** 14 pre-2023 wave-pool heats (Surf Ranch 2018, Freshwater 2019, Jeep Surf Ranch 2021) carry the sentinel value `-1` because WSL's leaderboard rendering for those years does not expose individual heat IDs. The verifier uses a leaderboard-format branch for these. No data is missing — only the WSL-side ID.
+
+5. **Source omissions in `wave_range` / `wind_condition`.** WSL did not consistently publish wave range or wind condition for 2014–2017 events, and a few 2023 events have event-wide gaps (Pipe Pro, Portugal, Bells Beach). NULL means absent at source. ERA5 covers these gaps numerically.
+
+6. **ERA5 break coverage = 28 of 40 breaks.** 98.5% of heats join. The 112 unjoined heats are at 12 breaks not yet pulled into `ERA5Conditions` — Surf Ranch, Surf Abu Dhabi, plus 10 lower-volume venues. Open-Meteo data is available for these; pulling it is straightforward future work.
+
+7. **Surfer height / weight ~40% populated.** WSL profile pages omit these fields for many historical surfers. NULL = not published, not zero.
+
+---
+
+## Suggested uses
+
+**Fantasy surf league optimization.** The dataset was originally built for this. Per-event point totals, per-surfer break-history aggregates, and ERA5 conditions are all queryable in SQL. See [GitHub repo / paper / blog post — TBD] for prior modeling experiments.
+
+**Educational ML.** A worked example of a domain-expert-driven ML project from data collection through audit. The v4 build itself is a story — the project includes documented bug discoveries, parser corrections, and audit-driven verification. Useful for teaching how to build trustworthy data pipelines from scraped web sources.
+
+**Sports analytics research.** Wave-by-wave granularity over 12 years, multi-tour, multi-venue, with weather covariates. Suitable for studying career trajectories, condition-performance interactions, head-to-head matchup outcomes, etc.
+
+**What this dataset is NOT.** It is not an official WSL product. It is not real-time. It does not include video, replays, or judge-by-judge score breakdown (WSL does not publish per-judge scores in popup HTML for completed heats). It does not include interference Penalty-1 halving annotations (see Limitations).
+
+---
+
+## Schema details
+
+Run `.schema` in `sqlite3` against `wsl-v4.sqlite` for the canonical schema. Highlights:
+
+**Wave columns:**
+- `wave_id` (PK), `heat_id`, `athlete_id`, `wave_index` (1-based per surfer-heat)
+- `score` (REAL) — WSL's displayed score, post-penalty if applicable. NULL for TBD waves.
+- `is_counted` (INTEGER 0/1) — 1 if this wave is one of the surfer's top-2 (counts toward heat total).
+- `is_zeroed_by_interference` (INTEGER 0/1) — 1 only for Penalty 2/3 strict zeroing (see Limitations).
+- `direction` (TEXT) — `"right"` or `"left"` for wave-pool waves; NULL for ocean.
+
+**Heat columns:**
+- `heat_id` (PK), `round_id`, `event_id`, `heat_number`
+- `heat_date` (TEXT YYYY-MM-DD), `heat_duration` (TEXT ISO-8601, e.g. `PT30M`)
+- `wave_range`, `wind_condition` (TEXT — WSL's rendered strings, may be NULL for archive gaps)
+- `break_id` (TEXT) — joins to `Break` and `ERA5Conditions`
+- `wsl_heat_id` (INTEGER) — sentinel-aware: `> 0` = popup-format with WSL-assigned ID; `-1` = leaderboard-format pre-2023 wave-pool; never NULL
+- `data_quality_notes` (TEXT) — non-NULL for ~35 heats with documented anomalies
+
+**Event columns:**
+- `event_id` (PK), `name`, `year`, `tour` (`"MCT"` or `"WCT"`), `location`
+- `break_id` (TEXT) — default break for the event; heat-level break_id is authoritative for split events
+- `contest_start_date`, `contest_end_date`, `is_split_event` (INTEGER 0/1)
+- `note` (TEXT) — descriptive notes for calendar-straddling, COVID-disrupted, multi-break sourcing, etc.
+
+---
+
+## Citation
+
+If you use this dataset in research, please cite:
+
+```bibtex
+@dataset{wsl_v4_2026,
+  author       = {Friedlander, Mark},
+  title        = {WSL Championship Tour Heat-Level Dataset (2014--2025), v4},
+  year         = {2026},
+  publisher    = {Hugging Face},
+  url          = {https://huggingface.co/datasets/<TBD>}
+}
+```
+
+---
+
+## License
+
+This dataset is released under **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)**.
+
+Required attributions:
+- **WSL data:** Scraped from public results pages on `worldsurfleague.com`. Not an official WSL product.
+- **ERA5 weather data:** "Weather data by [Open-Meteo.com](https://open-meteo.com)" (CC BY 4.0).
+
+---
+
+## Changelog
+
+- **v4 (2026-04-30, this release):** Fresh recrawl of all competition data from WSL. Hash chain locked. ERA5 join coverage 98.5%. 365/365 stratified-sample heats verified clean. Audit verdict HEALTHY.

wsl-v4.sqlite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cb09c4452410e9ea874a73c139fca7923129f9bd698d4a811d57f871b9eeb95
+size 19406848