Boardsesh climb generator models

Small models that write climbs for standardised LED training boards, for the experimental climb generator in the Boardsesh mobile app. The app downloads a board's files the first time you open the generator on that board, checks them against that generation's manifest.json, and keeps them on the phone.

There are two generations, each in its own folder:

  • v2/: every board below. A climb writer plus a grader.
  • v3/: Kilter Original and Tension Board 1, for now. It adds a "classic" quality token to the writer, a realism critic that ranks down climbs that don't look human-set, a reach safety net, and a check that rejects near-copies of existing climbs. The app has a v2/v3 switch on those boards so the two can be compared on the wall. Details: V3.md.

Training code: ml/climbgen.

Boards (v2)

Each folder in v2/ is one board layout, named <board>-<layout id> after the layout IDs in the board catalogue.

Folder Board Training rows Grader error Real grade inside the shown range, V0–V7 / V8–V11
kilter-1 Kilter Board Original (12x12) 126k 0.99 V 84–86% / 80%
tension-9 Tension Board 1 40k 0.96 V 82–86% / 79%
tension-10 Tension Board 2 Mirror 24k 1.05 V 84–86% / 79%
tension-11 Tension Board 2 Spray 19k 1.08 V 82% / 73%
kilter-8 Kilter Board Homewall (10x12) 24k 0.91 V 85% / 78%
moonboard-2 MoonBoard 2016 50k 1.09 V 90% / 80%
moonboard-3 MoonBoard 2024 18k 1.07 V 90–91% / 84%
moonboard-4 MoonBoard Masters 2017 49k 1.38 V 88–89% / 90%
moonboard-5 MoonBoard Masters 2019 33k 1.24 V 88% / 81%

The two Tension Board 2 layouts share most holes but bolt different holds into them, so each has its own model. The four MoonBoards are different hold sets on one 11x18 grid; their holds are numbered by grid cell, and they have no foot-only role.

The MoonBoard graders give wider ranges than the others (about 4.5–6 grades on Masters 2017, Masters 2019 and 2024, against about 3.5 elsewhere), and read the easiest climbs (6A+, about V3–V4) up to a grade high, because most MoonBoard climbs sit at the bottom of the scale.

Grader numbers are from climbs held out of training and tuning. "Grader error" is how far the grader's middle guess is from the climb's crowd grade, on average. The last column takes the range the app shows for a climb (for example "V8–V10") and counts how often the climb's crowd grade falls inside it. It's given separately for easier climbs (V0–V7) and harder ones (V8–V11), because hard climbs are rarer and the grader is less sure of them. The ranges aim for about 80%.

Files

v2/manifest.json                             every board's current version, file paths and SHA-256
v2/<board>/<version>/weights.bin             the climb writer: a small transformer, int8 weights
v2/<board>/<version>/grader.json             the grader: a probability for each V-grade, so the app can show a range
v2/<board>/<version>/model-meta.json         the board's holds and positions, hold roles, angles, keywords, tensor layout

v3/manifest.json                             the same, for v3
v3/<board>/<version>/weights.bin             the climb writer, trained with the quality token
v3/<board>/<version>/grader.json             the grader (v2's)
v3/<board>/<version>/critic.json             the realism critic, tied to its writer by SHA-256
v3/<board>/<version>/geometry.json           the longest reaches and gaps real classics have, per grade and angle
v3/<board>/<version>/known-climbs.bin        every known climb's holds, for rejecting near-copies
v3/<board>/<version>/model-meta.json         as in v2

Paths in each manifest are from the repo root (v2/..., v3/...).

A retrained model goes in a new version folder (a date) and its generation's manifest.json points at it; phones pick it up on their next open and delete the old one.

Data

Trained on the public nightly board snapshots Boardsesh publishes (dataset docs): listed climbs with at least 3 ascents, one example per climb and angle. The climbs and grades are created by each board's climbing community.

Kilter, Tension and MoonBoard are trademarks of their respective owners. These models and Boardsesh are not affiliated with or endorsed by Aurora Climbing, Moon Climbing or any board manufacturer.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support