README.md

---
license: cc-by-4.0
pretty_name: CONUS Flash-Flood Benchmark (L1-L3)
tags: [hydrology, flash-flood, benchmark, usgs, noaa, mrms, streamflow]
size_categories: [10K<n<100K]
configs:
- config_name: l1_episodes
  data_files: [{split: train, path: catalog/l1_episodes.parquet}]
- config_name: l1_locations
  data_files: [{split: train, path: catalog/l1_locations.parquet}]
- config_name: l2_episodes
  data_files: [{split: train, path: catalog/l2_episodes.parquet}]
- config_name: l3_testbeds
  data_files: [{split: train, path: catalog/l3_testbeds.parquet}]
- config_name: l3_hwms
  data_files: [{split: train, path: catalog/l3_hwms.parquet}]
- config_name: manifest
  data_files: [{split: train, path: manifest.parquet}]
---

# CONUS Flash-Flood Benchmark (L1-L3) — agent-friendly, CLI-queryable

Query by **level (L1/L2/L3) + US state + time range** and download from a bare terminal.
Three levels: **L1** national NCEI flash-flood episode catalog (1996-2025, 42,466), **L2**
observation-availability catalog (2021-2025, 5,424), **L3** evaluation-ready gauge testbeds
(688; 152 strict flash-floods; 47 NCEI-confirmed; 26 with return period >= 2 yr).

## TL;DR for agents — manifest = discover, jq = filter, resolve_url = download
```bash
BASE=https://huggingface.co/datasets/skyan1002/flash-flood-benchmark-data/resolve/main
curl -sL "$BASE/manifest.jsonl" -o manifest.jsonl   # one comma-safe index of every record + file
```
Every line is one JSON object: `record_type` (record|artifact), `level`, `ff_episode_id`,
`gage_id`, `year`, `month`, `begin_date` (YYYY-MM-DD), `primary_state_abbrev`,
`all_states_abbrev`, plus L3 fields (`benchmark_tier`, `flash_flood_class`,
`is_strict_flash_flood`, `selected`, `lp3_return_period_yr`, ...) and, for artifacts,
`artifact_kind`, `resolve_url`, `bytes`, `sha256`.

## Quickstart recipes
```bash
# 1) Screen: L3 strict flash floods in TX, July 2025
jq -c 'select(.record_type=="record" and .level=="L3" and .primary_state_abbrev=="TX"
  and .is_strict_flash_flood==true and .begin_date>="2025-07-01" and .begin_date<="2025-07-31")
  | {testbed_id, lp3_return_period_yr, peak_value}' manifest.jsonl

# 2) Download every file for one testbed (resolve URLs precomputed; -L follows the LFS/CDN redirect)
jq -r 'select(.testbed_id=="FF_2025_07_TX_ep002__08167000" and .record_type=="artifact").resolve_url' \
  manifest.jsonl | xargs -n1 curl -L -O

# 3) Just the hydrograph CSVs for the curated set (selected & RP>=2), budget by bytes first
jq -r 'select(.selected==true and .lp3_return_period_yr>=2 and .artifact_kind=="streamflow")
  | "\(.bytes)\t\(.resolve_url)"' manifest.jsonl

# 4) L1 episodes that touched TX in 2015 (any-touch via all_states_abbrev)
jq -c 'select(.level=="L1" and (.all_states_abbrev|test("(^|\\|)TX(\\||$)")) and .year==2015)
  | {ff_episode_id, primary_state_abbrev, begin_date}' manifest.jsonl

# 5) One-shot curated bundle (26 selected & RP>=2, ~0.2 GB)
curl -L "$BASE/bundles/testbeds_selected26.zip" -o selected26.zip

# 6) A single light file (no zip): one watershed polygon
curl -L "$BASE/testbeds/FF_2025_07_TX_ep002/08167000/watershed.geojson" -o ws.geojson
```

## Optional: server-side filter (no manifest download)
```bash
curl -sG https://datasets-server.huggingface.co/filter \
  --data-urlencode dataset=skyan1002/flash-flood-benchmark-data \
  --data-urlencode config=l3_testbeds --data-urlencode split=train \
  --data-urlencode "where=\"primary_state_abbrev\"='TX' AND year=2025 AND is_strict_flash_flood=true" \
  --data-urlencode length=100
```
Rules: double-quote column names, single-quote strings, URL-encode, length<=100 (page with offset),
re-index may lag minutes after an update.

## Contracts (read before scripting)
- **State:** filter on `primary_state_abbrev` (2-letter). For multi-state episodes use
  `all_states_abbrev` (pipe-joined, e.g. `NJ|NY|PA`).
- **Time:** `begin_date`/`end_date` are `YYYY-MM-DD` (lexical compare == date range); `year`/`month`
  are robust integer handles across levels.
- **Download:** always `curl -L` (LFS objects 302-redirect; without -L you get a tiny pointer). Use
  `-C -` to resume; check the `bytes` column before multi-GB pulls.
- **Never** `awk -F,` / `cut -d,` the raw CSVs in `catalog/raw_csv/` — fields contain commas. Use
  `jq` on `manifest.jsonl`, the Parquet configs, the `/filter` API, or a real CSV parser.
- **zarr** gridded forcing ships inside `packages/<id>.zip`; unzip then open with `zarr.open_group`
  (NOT `xr.open_zarr`). The light `mrms_2min_precip_basin_mean.csv` is the curl-friendly alternative.

## Helper CLI (optional)
`ffbench` (POSIX sh; needs curl, uses jq if present) wraps the same URLs:
```bash
curl -sL "$BASE/ffbench" -o ffbench && chmod +x ffbench
./ffbench query  --level l3 --state TX --start 2025-07-01 --end 2025-07-31 --strict
./ffbench get FF_2025_07_TX_ep002__08167000 --out ./tx_testbed
./ffbench bundle selected26
```

## Layout
`catalog/*.parquet` (+ `catalog/raw_csv/`), `manifest.{jsonl,csv,parquet}`, `manifest_schema.json`,
`testbeds/<ep>/<gage>/` (loose light files), `packages/<ep>__<gage>.zip` (full per-testbed),
`bundles/` (selected26, strict152). Figures/interactive map: companion Space
`skyan1002/flash-flood-benchmark` and dataset `skyan1002/flash-flood-benchmark-figures`.

## License & citation
CC-BY-4.0. Derived from public USGS (NWIS, STN, NLDI, 3DEP) and NOAA (NCEI Storm Events, NSSL
MRMS/FLASH-CREST) data. See `CITATION.cff`.