Quattro Formaggi: Local Parsing of Freight Requests with a Fine-Tuned Language Model

This paper describes Quattro Formaggi, an experimental pipeline that converts Russian and English freight requests into structured shipment cards. Mistral-Nemo-Instruct-2407 was fine-tuned with QLoRA on 1,500 synthetic examples and converted to GGUF for local inference. The evaluation set contains 271 requests: 238 template-generated records and 33 mock requests written by an AI assistant. The main metric is the proportion of cards with all key fields correct among 222 eligible records. In a Transformers comparison, fine-tuning raised it from 1.35% to 95.05%. The final Q6_K model in LM Studio reached 95.05%, against 87.39% for a rule-based extractor; on the 27 eligible mock requests, the scores were 74.1% and 70.4%. The main-metric threshold was met with no margin in the number of correct cards, but two criteria concerning shipping-condition errors were not met. The paper describes data construction, evaluation, and deployment to a local runtime. The results apply to the specified schema and synthetic set; performance on real requests was not measured.