Generate a climb
Predict grade
Click holds to build a climb
What the controls mean
--
-
- Temperature -
- Sampling randomness. Lower values are more conservative; higher values are more exploratory. - -
- Target V-grade -
- The grade token given to the generator. The generated route is also checked by the grade predictor. - -
- Known climb -
- An exact match against the tokenized dataset: same board, same angle, and same hold-role set. - -
- Validity -
- A structural check: enough holds, no duplicate placements, at least one start, and at least one finish. -
Research demo caveat
-- This is an experimental model demo. Generated climbs and predicted grades may be wrong, especially at rare grades or sparse board/angle combinations. -
-How this works
-
- Routes are converted into tokens such as
- <BOARD_TB2>, <ANGLE_40>,
- and <TB2_p652_start>.
-
- The generator samples a token sequence. The grade predictor removes the grade token and estimates difficulty from the board, angle, and hold-role tokens. -
-Links
- -Angle scope
-- The physical boards can be used at steeper angles than this demo exposes. This model snapshot is intentionally restricted to the angle range used in training/evaluation: TB2 up to 50° and Kilter up to 55°. -
-- The restriction avoids asking the models to extrapolate into sparse, noisier high-angle data where grades and route distributions are less reliable. -
-Data acknowledgement
-- Board layouts, hold metadata, and route data are derived from Tension Board 2 and Kilter Board datasets. - This project is unaffiliated with Tension Climbing or Kilter. - The route generator is inspired by Andrej Karpathy's - nanoGPT. -
-Choose a board and run a request
@@ -175,9 +104,61 @@What the controls mean
+-
+
- Temperature +
- Sampling randomness. Lower values are more conservative; higher values are more exploratory. +
- Target V-grade +
- The grade token given to the generator. The generated route is also checked by the grade predictor. +
- Known climb +
- An exact match against the tokenized dataset: same board, same angle, and same hold-role set. +
- Validity +
- A structural check: enough holds, no duplicate placements, at least one start, and at least one finish. +
Research demo caveat
+This is an experimental model demo. Generated climbs and predicted grades may be wrong, especially at rare grades or sparse board/angle combinations.
+How this works
+Routes are converted into tokens such as <BOARD_TB2>, <ANGLE_40>, and <TB2_p652_start>.
The generator samples a token sequence. The grade predictor removes the grade token and estimates difficulty from the board, angle, and hold-role tokens.
+Links
+ +Angle scope
+The physical boards can be used at steeper angles than this demo exposes. This model snapshot is intentionally restricted to the angle range used in training/evaluation: TB2 up to 50° and Kilter up to 55°.
+The restriction avoids asking the models to extrapolate into sparse, noisier high-angle data where grades and route distributions are less reliable.
+Data acknowledgement
+Board layouts, hold metadata, and route data are derived from Tension Board 2 and Kilter Board datasets. This project is unaffiliated with Tension Climbing or Kilter. The route generator is inspired by Andrej Karpathy's nanoGPT.
+