MikhailSAI/Quattro-Formaggi-12B-Logistics-v0.1

Quattro-Formaggi-12B-Logistics-v0.1: Specialized Logistics Data Extraction

This model, developed by MikhailSAI, is a 12 billion parameter variant derived from mistralai/Mistral-Nemo-Instruct-2407. It has been fine-tuned using a QLoRA adapter specifically for the task of transforming free-text freight requests (in Russian or English) into structured JSON shipment cards. The model identifies key logistics data such as route, dates, equipment, cargo details, and special conditions, while also flagging missing information and conflicting values.

Key Capabilities

• Structured Data Extraction: Converts natural language freight requests into a predefined JSON schema (card_v2).
• Multilingual Support: Processes requests in both Russian and English.
• Conflict and Missing Field Identification: Outputs lists of fields that are missing or contain conflicting values.
• Logistics-Specific Data: Extracts details like pickup/delivery locations, dates, cargo weight, equipment type, and special conditions (e.g., temperature, securing).

Training and Limitations

The model was trained exclusively on synthetic data derived from the yogape/logistics-operations dataset, with a focus on 20 Russian cities. While it performs well on template-based synthetic requests (achieving 0.9505 key-field accuracy on bench_v2), its performance on more informal, real-world-like requests (33 mock requests) is lower (0.741 key accuracy). It is a research release (v0.1) and is not intended for legal, compliance, or financial decisions. Crucially, unit conversions, total weight calculations, and the definitive missing-fields rule are handled by external code, not the model itself. Users must provide a specific system prompt and request date format for optimal performance.