Training Lab
Understand every move. Build a stronger model.
Board
White to moveEngine analysis
Search experiments
Candidate lines
Feature contributions
Weighted contributions before the score transform.
All features & search diagnostics
| Feature | Value | Weight | Contribution |
|---|
Position dataset
Import games, inspect coverage, and label positions with your local Stockfish.
No positions yet.
| Position | Source / group | Phase | Stockfish · White | Depth / nodes | Action |
|---|
Train an evaluator
Fit the 54-feature model to Stockfish scores. Whole games stay in one split; shared transpositions are excluded.
Mate and bound-only labels are excluded. Scores use a tanh target and the existing model's inverse transform.
Checkpoints are saved in this browser. A closed tab pauses training.
Learning curves
ReadyModel arena
Paired openings, reversed colours, equal node budgets. Compare candidates against a fixed baseline.
No comparison yet. Short matches check functionality; they do not establish an Elo gain.
| Game | Candidate colour | Result | Reason | Plies | Review |
|---|
Model versions
Inspect, edit, compare, and export local weights. The live ranked model is managed separately.
| Model | Created | Validation loss | Held-out error | Action |
|---|
Inspect and edit feature weights
Changes create a new local model. Contributions shown in Inspect use the active model.
| Feature | Weight | Change from deployed |
|---|
Suggest a model
Loading access…
Anyone can train and play locally. Sending a suggestion shares a copy of the weights and this note with the owner. Your dataset, games, and local Stockfish connection remain on your computer.
Guest suggestions are linked to this browser's session cookie. Sign in before submitting to keep access across browsers. Only the owner can publish. Training loss on different visitor datasets is not a fair comparison; the server runs a fixed shared test suite.