Dobrovolskyi, I. Security Document Classification with a Fine-Tuned Local Large Language Model: Benchmark Data and an Open-Source System. Journal of Information Security and Applications, 2026.

Benchmark results

Evaluated under identical methodology (alpaca prompt, Ollama /api/generate, temperature = 0, num_predict = 2048) on the companion dataset torchsight/cybersecurity-classification-benchmark. Canonical numbers live in that repo's BENCHMARK_NUMBERS.md.

Primary — eval-1000-synthetic (n = 1,000)

Model	Type	Cat. acc [95% CI]	Subcat. acc
Beam q4_K_M	Local (LoRA)	95.0% [93.5, 96.2]	48.2%
Beam f16	Local (LoRA)	93.2% [91.5, 94.6]	51.1%
Beam q8_0	Local (LoRA)	93.0% [91.2, 94.4]	51.4%
Claude Sonnet 4	Commercial API	79.9% [77.3, 82.3]	23.0%
Claude Opus 4	Commercial API	79.9% [77.3, 82.3]	22.5%
GPT-5	Commercial API	76.9% [74.2, 79.4]	11.6%
Gemini 2.5 Pro	Commercial API	75.4% [72.6, 78.0]	21.0%
Qwen 3.5 27B base	Local (no LoRA)	86.3% [84.0, 88.3]	19.0%
Regex (48 patterns)	Rule-based	52.7% [49.6, 55.8]	—

q8_0 achieves the highest subcategory accuracy (51.4%) of the three Beam variants but slightly lower category-level accuracy than q4_K_M.

External — eval-500-external (n = 500)

Model	Cat. acc [95% CI]	Δ vs. primary
Beam q4_K_M	93.8% [91.3, 95.6]	−1.2 pp
Beam f16	91.2% [88.4, 93.4]	−2.0 pp
Beam q8_0	91.2% [88.4, 93.4]	−1.8 pp
Claude Sonnet 4	86.4% [83.1, 89.1]	+6.5 pp
Gemini 2.5 Pro	82.0% [78.4, 85.1]	+6.6 pp
Qwen 3.5 27B base	86.6% [83.3, 89.3]	+0.3 pp
GPT-5	65.8% [61.5, 69.8]	−11.1 pp
Regex baseline	29.6% [25.8, 33.7]	−23.1 pp

Usage with Ollama

bash

ollama pull torchsight/beam-q8_0
ollama run torchsight/beam-q8_0

Or via the TorchSight CLI.

Training

Base: Qwen 3.5 27B (dense)
Method: LoRA (r = 128, α = 256), bf16, 5 epochs
Dataset: 78,358 balanced samples — see torchsight/beam-training-data
Hardware: 8× NVIDIA A100 80GB SXM4, 10.5 hours

License

Apache 2.0. The base model (Qwen 3.5 27B) carries its own license; consult upstream terms for use.