machina-sports/ayrton-1

Ayrton-1: Formula 1 Domain-Specialist LLM

Ayrton-1 is a 12 billion parameter language model developed by Machina Sports, fine-tuned from google/gemma-3-12b-it. It is specifically engineered to provide factual answers to Formula 1-related questions, covering the 1950–2025 seasons. The model is named after Ayrton Senna and is optimized for local inference on Apple Silicon using MLX.

Key Capabilities

• Factual F1 QA: Answers questions on race results, championship standings, and driver/constructor history from 1950–2025.
• Session-level Data: Provides details on telemetry and strategy for the 2018–2025 seasons, leveraging data from FastF1.
• High Accuracy: Achieves hybrid accuracy scores of 0.957 on 2025 test data and 0.945 on 2024 validation data for value-level matches with paraphrase tolerance.
• Historical Reasoning: Functions as a frozen knowledge model, not a live-data source or general-purpose chatbot.

Training and Limitations

The model was trained using iterative LoRA fine-tuning on Apple Silicon, utilizing the machina-sports/ayrton-1-qa-v2 dataset. It was distilled using Gemini 3 Flash for style and fluency. Limitations include imperfect refusal calibration for pre-2018 telemetry, a coverage skew towards the modern era (2018–2025), and an inability to provide real-time information or data beyond the 2025 season. It is English-only and not intended for general chat, code, or math tasks.

Good For

• Developers needing a specialized model for historical Formula 1 data retrieval.
• Applications requiring factual F1 information for specific seasons and events.
• Local inference on Apple Silicon devices.