GenClimb: AI-Generated Climbing Routes for Interactive Training Boards

GenClimb is a generative AI model designed to create climbing routes for Standardized Interactive Climbing Training Boards (SICTBs). It a seq2seq transformer architecture, GenClimb generates climbs based on board layouts and climb difficulties.

Model Details

Architecture

  • Dimension: 512
  • Attention Heads: 4
  • Layers: 5
  • Feed-Forward Dimension: 1024
  • Dropout Rate: 0.15
  • Activation Function: GELU
  • Layer Normalization Epsilon: 1e-5

Training Configuration

  • Device: CUDA (1x NVIDIA GeForce RTX 3070)
  • Learning Rate: 1e-4
  • Epochs: 8
  • Weight Decay: 0.0125
  • Batch Size: 32
  • Train/Test Split: 90/10

Performance Metrics

  • Training Time: 12 hours and 6 minutes
  • Final Loss: 2.114803

Dataset

The model is trained on the Kilter-Board-Dataset, a comprehensive collection of climbing routes curated with the help of the BoardLib utility by lemeryfertitta.

Downloads last month
5
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train stfamod/genclimb-large-quantized