docs(cards): 70 Grand Prix, canonical lap-time MAE, and the real train/test split

README.md

@@ -1,68 +1,72 @@
----
-license: apache-2.0
-language:
-- en
-tags:
-- formula-1
-- motorsport
-- race-strategy
-- multi-agent
-- langgraph
-- xgboost
-- lightgbm
-- temporal-convolutional-network
-- sports-analytics
----
-
-# F1 StratLab Strategy Models
-
-The machine learning models behind F1 StratLab, an open-source multi-agent system for
-Formula 1 race strategy. Six LangGraph sub-agents and a ReAct orchestrator call these
-models to produce pit-stop recommendations, tire-degradation forecasts, overtake and
-undercut probabilities, and answers grounded in the FIA regulations.
-
-Links:
-
-- Project: https://f1stratlab.com
-- Documentation: https://docs.f1stratlab.com
-- Source code: https://github.com/VforVitorio/F1-StratLab
-- Dataset: https://huggingface.co/datasets/VforVitorio/f1-strategy-dataset
-
-## Models
-
-| Task | Algorithm | Metric |
-|---|---|---|
-| Lap-time prediction | XGBoost | MAE 0.392 s |
-| Tire degradation | TCN with Monte Carlo Dropout | P10/P50/P90 quantiles, pit-window detection |
-| Overtake probability | LightGBM | AUC-ROC 0.876 |
-| Safety-car probability | LightGBM | classifier |
-| Pit-stop duration | HistGradientBoosting (quantile) | MAE 0.487 s |
-| Undercut success | LightGBM (binary) | AUC-ROC 0.771 |
-| Team-radio NLP | Whisper, RoBERTa, SetFit, BERT-large | 4-stage pipeline |
-
-## Training data
-
-Trained on telemetry, lap data and race-control messages from 71 Grand Prix across the
-2023 to 2025 seasons, taken from the FastF1 and OpenF1 public APIs. The processed data is
-published as a companion dataset: VforVitorio/f1-strategy-dataset.
-
-## Intended use
-
-Research and educational use for Formula 1 strategy analysis. Not affiliated with Formula 1,
-the FIA or any team. Predictions are estimates, not guarantees.
-
-## Citation
-
-```bibtex
-@misc{vegasobral2026f1stratlab,
-  author = {Vega, V{\'i}ctor},
-  title  = {F1 StratLab: AI Models for Strategy Recommendations in Formula 1 Races},
-  year   = {2026},
-  note   = {Bachelor's Thesis, Intelligent Systems Engineering, UIE Campus Coru{\~n}a},
-  url    = {https://f1stratlab.com}
-}
-```
-
-## License
-
-Apache 2.0. Author: Víctor Vega (https://github.com/VforVitorio).
+---
+license: apache-2.0
+language:
+- en
+tags:
+- formula-1
+- motorsport
+- race-strategy
+- multi-agent
+- langgraph
+- xgboost
+- lightgbm
+- temporal-convolutional-network
+- sports-analytics
+---
+
+# F1 StratLab Strategy Models
+
+The machine learning models behind F1 StratLab, an open-source multi-agent system for
+Formula 1 race strategy. Six LangGraph sub-agents and a ReAct orchestrator call these
+models to produce pit-stop recommendations, tire-degradation forecasts, overtake and
+undercut probabilities, and answers grounded in the FIA regulations.
+
+Links:
+
+- Project: https://f1stratlab.com
+- Documentation: https://docs.f1stratlab.com
+- Source code: https://github.com/VforVitorio/F1-StratLab
+- Dataset: https://huggingface.co/datasets/VforVitorio/f1-strategy-dataset
+
+## Models
+
+| Task | Algorithm | Metric |
+|---|---|---|
+| Lap-time prediction | XGBoost | MAE 0.410 s |
+| Tire degradation | TCN with Monte Carlo Dropout | P10/P50/P90 quantiles, pit-window detection |
+| Overtake probability | LightGBM | AUC-ROC 0.876 |
+| Safety-car probability | LightGBM | classifier |
+| Pit-stop duration | HistGradientBoosting (quantile) | MAE 0.487 s |
+| Undercut success | LightGBM (binary) | AUC-ROC 0.771 |
+| Team-radio NLP | Whisper, RoBERTa, SetFit, BERT-large | 4-stage pipeline |
+
+## Training data
+
+Built from telemetry, lap data and race-control messages covering 70 Grand Prix across the
+2023 to 2025 seasons, taken from the FastF1 and OpenF1 public APIs. The seasons are not
+interchangeable: every model **trains on 2023 and 2024** and is **tested on 2025**, which is
+the holdout the shipped system infers on (`train_seasons: [2023, 2024]`, `test_season: 2025`
+in each model config). Quoting a figure that pools all three is partly the system reading
+back its own training data. The processed data is
+published as a companion dataset: VforVitorio/f1-strategy-dataset.
+
+## Intended use
+
+Research and educational use for Formula 1 strategy analysis. Not affiliated with Formula 1,
+the FIA or any team. Predictions are estimates, not guarantees.
+
+## Citation
+
+```bibtex
+@misc{vegasobral2026f1stratlab,
+  author = {Vega, V{\'i}ctor},
+  title  = {F1 StratLab: AI Models for Strategy Recommendations in Formula 1 Races},
+  year   = {2026},
+  note   = {Bachelor's Thesis, Intelligent Systems Engineering, UIE Campus Coru{\~n}a},
+  url    = {https://f1stratlab.com}
+}
+```
+
+## License
+
+Apache 2.0. Author: Víctor Vega (https://github.com/VforVitorio).