---
  title: TorchSight
  emoji: 🔥
  colorFrom: orange
  colorTo: red
  sdk: static
  pinned: false
  ---

  # TorchSight

  On-premises cybersecurity scanner powered by fine-tuned LLMs. Detects credentials, PII, malicious payloads, and sensitive data in text, images, and PDFs — without sending data to the cloud.

  ## Models

  - [beam-q4_K_M](https://huggingface.co/torchsight/beam-q4_K_M) — 95.1% accuracy, 17GB (recommended)
  - [beam-q8_0](https://huggingface.co/torchsight/beam-q8_0) — 92.7% accuracy, 28GB
  - [beam-f16](https://huggingface.co/torchsight/beam-f16) — 93.0% accuracy, 53GB

  ## Datasets

  - [cybersecurity-classification-benchmark](https://huggingface.co/datasets/torchsight/cybersecurity-classification-benchmark) — 1052-sample eval benchmark
  - [beam-training-data](https://huggingface.co/datasets/torchsight/beam-training-data) — 74K SFT training samples

  ## Highlights

  - Beats Claude Opus 4 (79.9%), Gemini 2.5 Pro (75.4%) on domain classification
  - LoRA fine-tune of Qwen 3.5 27B, runs locally via Ollama
  - Apache 2.0 licensed

  [GitHub](https://github.com/IvanDobrovolsky/torchsight)