skyan1002's picture
v1.0 paper-matched: 806 L3 / 21 primary, agent flags, master_metadata, AGENTS.md + tool schema, annual L3-expansion pipeline + tools
7260f53 verified
Raw
History Blame Contribute Delete
3.95 kB
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"dataset": "skyan1002/flash-flood-benchmark-data",
"base_url": "https://huggingface.co/datasets/skyan1002/flash-flood-benchmark-data/resolve/main",
"description": "LLM/agent tool definitions for the TORRENT flash-flood benchmark. Each tool maps to a jq query over manifest.jsonl or the /filter API; see AGENTS.md.",
"tools": [
{
"name": "list_primary_benchmark",
"description": "Return the 21-testbed conservative primary benchmark (14 episodes, 21 gauges) — the default set for model intercomparison.",
"parameters": {"type": "object", "properties": {}, "required": []},
"implementation": "jq -c 'select(.level==\"L3\" and .primary_benchmark==true)' manifest.jsonl"
},
{
"name": "query_testbeds",
"description": "Filter L3 testbeds by tier of evidence, flash-flood screen, state, time, and severity. Returns matching testbed records.",
"parameters": {
"type": "object",
"properties": {
"tier": {"type": "string", "enum": ["primary", "rp_ge_2yr", "ncei_confirmed", "dedup", "strict", "all"], "description": "which stage of the funnel"},
"state": {"type": "string", "description": "2-letter USPS code, matches primary_state_abbrev"},
"start_date": {"type": "string", "description": "YYYY-MM-DD inclusive lower bound on begin_date"},
"end_date": {"type": "string", "description": "YYYY-MM-DD inclusive upper bound on begin_date"},
"min_return_period_yr": {"type": "number", "description": "lower bound on lp3_return_period_yr"},
"max_drainage_area_km2": {"type": "number"}
},
"required": []
},
"implementation": "jq over manifest.jsonl with level=='L3' plus the flag matching the tier (primary->primary_benchmark, rp_ge_2yr->rp_ge_2yr, ncei_confirmed->ncei_confirmed, dedup->selected, strict->is_strict_flash_flood)"
},
{
"name": "get_testbed",
"description": "List and download every file of one testbed package (watershed.geojson, streamflow.csv, hwms_inside.csv, testbed_summary.json, and full forcing if present).",
"parameters": {
"type": "object",
"properties": {"testbed_id": {"type": "string", "description": "e.g. FF_2025_07_TX_ep002__08165500"}},
"required": ["testbed_id"]
},
"implementation": "jq -r 'select(.testbed_id==$id and .record_type==\"artifact\").resolve_url' manifest.jsonl | xargs -n1 curl -L -O"
},
{
"name": "expand_l3",
"description": "Discover newly qualifying testbeds from a given NCEI calendar year and append them to the L3 pool as light rows (needs_curation_review=true). Extends the benchmark; does not alter the human-reviewed primary set.",
"parameters": {
"type": "object",
"properties": {"year": {"type": "integer", "description": "NCEI calendar year to ingest, e.g. 2026"}},
"required": ["year"]
},
"implementation": "python pipeline/expand_l3.py --year <year> --data-root ."
}
],
"flag_semantics": {
"primary_benchmark": "default 21-testbed conservative set (paper Sect. 3.4)",
"is_strict_flash_flood": "passed strict flash-flood screen (175)",
"selected": "deduplicated representative (115); alias dedup_selected",
"ncei_confirmed": "in-basin NCEI report point (54)",
"rp_ge_2yr": "LP3 return period >= 2 yr (31)",
"nldi_recovered": "watershed restored by hardened NLDI retrieval",
"regulation_excluded": "excluded from primary: direct flood-control influence",
"postfire_excluded": "excluded from primary: recent-wildfire applicability review",
"needs_curation_review": "machine-appended by the annual pipeline, awaiting human curation",
"water_balance_flag": "RC > 1; interpret volume metrics with caution",
"benchmark_tier": "A (in-basin NCEI point) or B (evidence label, not an exclusion)"
}
}