Grid Dispatch v1: 500 hard tasks (453 train / 47 test, split by day)

README.md

@@ -0,0 +1,113 @@
+---
+license: cc-by-4.0
+pretty_name: Grid Dispatch tasks
+tags:
+- reinforcement-learning
+- rl-environment
+- agents
+- tool-use
+- operations-research
+- power-systems
+- unit-commitment
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train.jsonl
+  - split: test
+    path: data/test.jsonl
+---
+
+# Grid Dispatch tasks
+
+500 tasks for the [Grid Dispatch](https://github.com/savarbhasin/grid-dispatch) RL environment (Verifiers v1). **An AI
+runs a power grid for one real day:** it decides which of 17 power plants are on each hour, and when to use a battery,
+to keep the lights on at the lowest cost.
+
+The day plays out in 4-hour steps. At each step the AI looks, tests ideas with `simulate`, gives orders, then advances
+the clock. The **real** day happens as it goes: demand differs from the forecast, wind falls short, and plants break
+down. Decisions can't be undone, and big plants need hours of notice, so a good dispatcher has to act before the problem
+arrives.
+
+| Split | Tasks | Real days |
+| --- | --- | --- |
+| `train` | 453 | 252 |
+| `test` | 47 | 28 |
+
+The split is **by real day**: one day can appear several times with a different fleet and breakdowns, and no day is in
+both splits. Test rows don't include the solver's moves.
+
+```python
+from datasets import load_dataset
+test = load_dataset("dev-front/grid-dispatch-tasks", split="test")
+```
+
+## How a task is made
+
+1. **A real day:** one 2024 day for a US grid area (SPP, ERCOT, MISO, PJM or ISNE) from EIA-930, giving hourly demand,
+   its real day-ahead forecast, and wind and solar output. Days where actual net load beat the forecast most are
+   preferred.
+2. **A fleet:** 17 plants sampled from the RTS-GMLC test system in PGLib-UC (nuclear, coal, combined cycle, steam,
+   peakers), plus a battery. The day is scaled so its peak fills 85% of the fleet.
+3. **Breakdowns:** 0–2 big plants trip during the day, mostly in the hours before the evening peak.
+4. **A reference dispatcher:** a solver (HiGHS) that re-plans every 4 hours with exactly the information the AI has,
+   keeping spare power for one breakdown or any two. Each day uses whichever version did better.
+5. **Keep only hard days:** following the midnight plan, and doing nothing, must both score ≤ 0.6 (on average they
+   score 0.10 and 0.15), and the reference must lose at most 1% of the day's power (it loses none on 333 of 500).
+
+Replaying the reference's moves reproduces its exact cost on every task.
+
+## Fields
+
+| Field | Meaning |
+| --- | --- |
+| `id` | `<area>-<date>-c<candidate>` |
+| `world.units` | The 17 plants: `pmin`, `pmax` (MW), `ramp`, `min_up` / `min_down` (hours), `lead` (hours of notice to start), `startup_cost`, `curve` (MW, $/h) |
+| `world.battery` | `power` (MW), `energy` (MWh), `eff` (one-way efficiency), `soc0` (starting charge share) |
+| `world.demand`, `world.demand_da` | Actual and day-ahead forecast demand per hour (MW, scaled) |
+| `world.renew`, `world.renew_da` | Actual wind+solar, and its forecast, per hour (MW, scaled) |
+| `world.trips` | Plant → hour it breaks. **Hidden from the AI** until it happens |
+| `world.init_on`, `world.init_p` | Which plants are on at midnight, and their output |
+| `reference.policy` | `rolling` (spare power for one breakdown) or `rolling2` (for any two) |
+| `reference.cost`, `reference.shed_mwh`, `reference.demand_mwh` | The reference's total cost, unserved energy, and the day's demand: what the score is measured against |
+| `reference.baselines` | Every policy played on this day (`rolling`, `rolling2`, `hindsight`, `static`, `myopic`, `nothing`) with its cost, unserved MWh, start-ups and score |
+| `reference.actions` | The reference's moves step by step (train only) |
+
+**`world` holds the hidden truth** (the actual demand, wind and trips). The environment only reveals it as the clock
+passes. Keep it out of prompts.
+
+## Score
+
+**Score = reliability × economy**, each against the reference:
+
+- **reliability** = `e^(−excess unserved MWh / (0.5% of the day's demand))`;
+- **economy** = `0.2 + 0.8·e^(−overspend / 10%)`, where overspend is fuel, no-load and start-up cost above the
+  reference's.
+
+Blackouts can't be bought off with fuel savings, and matching or beating the reference scores 1.0.
+
+## Baselines (the 5 hardest days, all from `train`; one attempt, no thinking cap)
+
+| Model | Average score |
+| --- | --- |
+| DeepSeek V4.1 Flash | 0.39 (one day 1.00, one 0.64, three ≤ 0.15) |
+| Nemotron Lightning 3.5 30B-A3B | 0.12 |
+
+DeepSeek never overspends, but on hard days it keeps too little spare power and is caught by the second breakdown.
+
+## Licence and attribution
+
+Released under **CC BY 4.0**. Built from:
+- **EIA-930** Hourly Electric Grid Monitor (US Energy Information Administration; US government data). In the 2024
+  files, PJM's forecast was shifted back one hour.
+- **[PGLib-UC](https://github.com/power-grid-lib/pglib-uc)** (CC BY 4.0), RTS-GMLC test system
+  ([GridMod/RTS-GMLC](https://github.com/GridMod/RTS-GMLC)). Changes: units sampled, renamed and rounded, and given
+  synthetic start-up notice times.
+
+**Synthetic parts:** start-up notice times, the wind/solar forecast (actual plus half of yesterday's error), the trips,
+the battery, and the scaling.
+
+The reward pattern (real records plus a solver reference) is inspired by
+[PortSimEnv](https://github.com/adithya-s-k/FineEnvs/tree/main/07-simulation-environments).