Shoreline

Land or water.

Type a latitude and longitude, or click the map. The tile is drawn from public-domain coastlines. The word beside it is the adapter’s answer for the one-degree cell that contains this place.

Model—
Atlas—
Score—

Drawing the atlas…

An episode, playing on its own

Reset, then an action, then a reward. Coast walks to the shore. Survey answers the cell. The loop is what the OpenEnv server steps through.

    Observation → Action → Reward

    Teaching a small model the shape of the coast

    Shoreline asks a plain question. Give a model a latitude and a longitude, and ask whether that place is land or water. The answer is one word. The score is whether that word matches an atlas.

    The atlas is a grid made from Natural Earth 1:110m coastlines and lakes, which are in the public domain. Each cell is a quarter of a degree. A harbor on that simplified coast can fall in the neighboring cell, so the label is the atlas, not a navigational chart. Lakes such as the Caspian Sea and the Great Lakes are water.

    The episode

    A survey episode is one step.

    Land or Water?
Latitude: 48.86
Longitude: 2.35
Reply with one word: land or water.

    The label is not in the prompt. Reward is 1 when the last land or water in the reply matches the cell, and 0 otherwise. Training draws land and water equally. On a uniform sample of the globe, water is the larger share, so a model that always says water looks better than it is.

    Two other episodes sit on the same atlas. Coast starts inland and asks the model to walk north, south, east, or west until it says here on the shoreline. Locate hides the coordinate, shows a coarse sketch of the view, and scores a guessed latitude and longitude by distance.

    Training

    The published adapter is a LoRA on Qwen/Qwen2.5-0.5B-Instruct, rank 8, on a single T4. Each step shows eight coordinates, half land and half water in expectation, and the target reply is the one word that would score 1. Four hundred steps cover 3,200 places. The weights are the adapter at vanishingradient/shoreline-land-water.

    47%Before training, on 200 balanced places. Land was almost never named.
    75.5%After training. 61% of land cells, 92% of water cells.
    81.5%On 200 places drawn uniformly across the globe.

    That is a coarse map, not a coastline survey. The model has a better chance on open ocean and on broad interiors than on a cell that sits on the 1:110m shore.

    On this page

    The map above takes a latitude and a longitude, or a click. It draws the atlas tile and shows the word the adapter gave for the one-degree cell that contains that place, next to the atlas label.

    The code is at github.com/saqlain2204/shoreline.

    python -m shoreline.demo --lat 48.86 --lon 2.35 --answer land