metadata
license: mit
pretty_name: FloodGate v1
tags:
- rl-environment
- openenv
- flood-control
- water-resources
- operations-research
- benchmark
- beam
configs:
- config_name: tasks
data_files: tasks/flood-v1-sample/tasks.jsonl
FloodGate v1: Cumberland River flood control on the real 2025 floods
Tasks, data and every model run for FloodGate. The task: run the U.S. Army Corps' four flood-control dams above Nashville through the February and April 2025 floods. The code is MIT. All data are U.S. federal works in the public domain (USACE, USGS, NOAA NWS); NWS products come via the Iowa Environmental Mesonet archive.
| path | what |
|---|---|
tasks/flood-v1-sample/tasks.jsonl |
24 tasks around 13-16 February and 1-4 April 2025, in three tiers: the record, a storm 1.5 times bigger, lakes starting half full. Each holds the dams, gauges, inflows and local runoff per 6-hour step, the travel times, and the reference (LP optimum and its plan, baselines, the Corps' real releases) |
data/basin.json |
the dams' official pool levels and release limits, the gauges' NWS flood categories and ratings, the travel times |
data/record.csv.gz |
the curated record, Jan-Jun 2025, 6-hour steps: dam releases, storage and pool elevations, gauge flows |
data/forecasts.csv.gz |
the 9,540 river-forecast values the NWS Ohio River Forecast Center issued for Burkesville, Celina and Nashville in Jan-Jun 2025 (stage, and flow through the NWS rating) |
data/SOURCE.md |
every rule and number, as real, derived or assumed |
runs/<run>/index.json |
every episode: reward, grade, the submitted plan (plan mode) or the commits (live mode), tokens, time |
runs/<run>/board.md |
the run's board |
runs/live-baselines-24h.md |
live mode's policies without a model: scenario-optimal 1.000, a perfect solver on the mean forecast 0.552, one scenario 0.444, hold 0.215 |
Results (live mode, 24 tasks)
| model | mean reward |
|---|---|
| a perfect solver planning on the mean forecast (no model) | 0.552 |
| DeepSeek-V4.1-Flash | 0.364 |
| Beam-1.0-501B-A23B (default effort) | 0.236 |
| Beam-1.0-501B-A23B (reasoning_effort high) | 0.237 |
| hold every release | 0.215 |
Grading
- Plan mode: reward = 0.2 + 0.8·exp(−ρ/0.25), where ρ = (cost − optimum) / (lazy − optimum). Plans that break a rule score below 0.2.
- Live mode: ρ = Σ regret / Σ (the lazy operator's regret). Each decision's regret is its expected extra cost over the scenario-optimal decision, from a two-stage stochastic LP over the 8 scenarios.
- Both are deterministic, with no LLM judge.
