How sogliuzzol/chessbehaviors works
ChessBehaviors is a research workflow for modeling and evaluating human chess move choices with Maia-2 and adaptation/search variants. Researchers build position datasets from PGNs, generate predictions (including empirical ground truth), evaluate accuracy and distribution/style divergence, then export tables and heatmaps. The graph preserves the documented capability branches and the core model, search, pruning, and style-representation layers. Edges use README workflow statements or indexed source references; unsampled implementation details are not inferred.
Main entry points: evaluate_style.py Style evaluation · base.py Policy wrappers · maia2_descent.py Search variants · evaluate_style_common_fens.py Common-FEN style
Components 33
Dataset preparation
- Dataset builder
build_dataset.py - Data utilities
utils/data.py - Position datasets
Prediction models
- Ground truth
ground_truth.py - Maia baseline
maia2_baseline.py - Fine-tuned Maia
maia2_ft.py - Nucleus variants
maia2_nucleus.py - Search variants
maia2_descent.py - MCTS variants
maia2_mcts.py - MoE-LoRA variants
maia2_moe_lora.py - Prediction files
Evaluation and analysis
- Move accuracy
evaluate_move_accuracy.py - CP error
evaluate_cp_error.py - JSD evaluation
evaluate_jsd.py - Common-FEN JSD
evaluate_common_fens_jsd.py - Style evaluation
evaluate_style.py - Common-FEN style
evaluate_style_common_fens.py - JSD heatmaps
generate_jsd_heatmaps.py - Style heatmaps
generate_style_heatmaps.py - Common-FEN heatmaps
generate_style_heatmaps_common_fens.py - LaTeX tables
tabs.py
Model internals
- Policy wrappers
models/base.py - Player fine-tuning
models/maia_ft.py - MoE-LoRA adapter
models/adapters/moe_lora.py - Training dataset
models/dataset.py - Player embeddings
models/embeddings.py - Policy pruning
utils/pruning.py - Descent search
search/descent.py - MCTS search
search/mcts.py - Board transitions
utils/transitions.py - Style manifold
models/style.py
Other components
- Researcher
- Raw PGN files
Connections 47
- Researcher → Dataset builder: builds data
README.md - Raw PGN files → Dataset builder: provides games
README.md - Dataset builder → Data utilities: uses utilities
build_dataset.py - Dataset builder → Position datasets: writes datasets
README.md - Position datasets → Ground truth: inputs positions
README.md - Position datasets → Maia baseline: inputs positions
README.md - Position datasets → Fine-tuned Maia: inputs positions
README.md - Position datasets → Nucleus variants: inputs positions
README.md - Position datasets → Search variants: inputs positions
README.md - Position datasets → MCTS variants: inputs positions
README.md - Position datasets → MoE-LoRA variants: inputs positions
README.md - Ground truth → Prediction files: writes predictions
README.md - Maia baseline → Prediction files: writes predictions
README.md - Fine-tuned Maia → Prediction files: writes predictions
README.md - Nucleus variants → Prediction files: writes predictions
README.md - Search variants → Prediction files: writes predictions
README.md - MCTS variants → Prediction files: writes predictions
README.md - MoE-LoRA variants → Prediction files: writes predictions
README.md - Prediction files → Move accuracy: evaluates moves
README.md - Prediction files → CP error: evaluates error
README.md - Prediction files → JSD evaluation: evaluates distributions
README.md - Prediction files → Common-FEN JSD: compares common FENs
README.md - Prediction files → Style evaluation: evaluates style
README.md - Prediction files → Common-FEN style: compares common FENs
README.md - JSD evaluation → JSD heatmaps: visualizes metrics
README.md - Style evaluation → Style heatmaps: visualizes metrics
README.md - Common-FEN style → Common-FEN heatmaps: visualizes metrics
README.md - Move accuracy → LaTeX tables: tabulates results
README.md - CP error → LaTeX tables: tabulates results
README.md - JSD evaluation → LaTeX tables: tabulates results
README.md - Style evaluation → LaTeX tables: tabulates results
README.md - Fine-tuned Maia → Player fine-tuning: uses model API
maia2_ft.py - Search variants → Policy wrappers: uses wrappers
maia2_descent.py - MCTS variants → Policy wrappers: uses wrappers
maia2_mcts.py - Search variants → Policy pruning: uses pruning
- MoE-LoRA variants → MoE-LoRA adapter: uses adapter
- Player fine-tuning → Training dataset: trains on
models/maia_ft.py - Player fine-tuning → Player embeddings: uses embeddings
models/maia_ft.py - MoE-LoRA adapter → Policy wrappers: wraps backbone
models/adapters/moe_lora.py - Policy wrappers → Descent search: delegates search
models/base.py - Policy wrappers → MCTS search: delegates search
models/base.py - Style evaluation → Board transitions: extracts transitions
evaluate_style.py - Style evaluation → Style manifold: uses style analysis
evaluate_style.py - Common-FEN style → Board transitions: extracts transitions
evaluate_style_common_fens.py - Common-FEN style → Style manifold: uses style analysis
evaluate_style_common_fens.py - Style heatmaps → Style manifold: uses style analysis
generate_style_heatmaps.py - Common-FEN heatmaps → Style manifold: uses style analysis
generate_style_heatmaps_common_fens.py