Security Document Classification with a Fine-Tuned Local Large Language Model: Benchmark Data and an Open-Source System
Organizations needing to classify sensitive documents locally without sending data to cloud services can benefit from this accurate, open-source alternative to commercial systems.
The paper introduces TorchSight, an open-source local system for security document classification using a fine-tuned Qwen 3.5 27B model, achieving 95.0% category-level accuracy on a benchmark of 1,000 documents, outperforming commercial models (75.4-79.9%) under the same protocol.
Organizations that scan documents for sensitive information face a practical problem. Cloud services require data to be sent to external infrastructure, while rule-based tools often miss threats that depend on context. This study presents TorchSight, an open-source local system for security document classification built around a fine-tuned Qwen 3.5 27B model. The model was trained on 78,358 samples from 13 permissively licensed sources and GPT-4 synthetic data covering seven security categories and 51 subcategories. In the main evaluation on 1,000 documents, the model reached 95.0% category-level accuracy (95% confidence interval: 93.5-96.2). The tested commercial models scored 75.4-79.9% under the same prompting protocol. On a separate external set of 500 held-out samples, the model reached 93.8% accuracy, which suggests that performance extends beyond the main benchmark, although the margin depends on dataset composition and difficult boundary cases. The results show that a fine-tuned local model can support accurate security document classification while keeping document processing under local control.