Acknowledgments
fsi-anomaly is original architecture and original training, built start-to-end on commodity hardware by Ferrell Synthetic Intelligence (FSI) — a solo, unfunded developer. This page exists because transparency is the product: we tell you exactly what was used to make this model, and what is original work.
Distillation teachers (training data)
The training gold for fsi-anomaly was authored and curated with heavy help from DeepSeek (V4) as the primary knowledge-distillation teacher, alongside Qwen, Kimi, and GPT-5.5 as assisting teachers.
Every example was hand-written, reviewed, and verified by FSI. The teacher models helped draft and refine the gold training set; the architecture, weights, training pipeline, and evaluation are original to FSI.
Infrastructure
- PyTorch — training and inference framework.
- Hugging Face ecosystem — tokenizers, model cards, repo hosting.
- llama.cpp — GGUF quantization and on-device inference.
- Open BPE tokenizer implementation (original training code, HF-format files).
The spirit of it
Built alone, on an 8-core ARM tablet, with no GPU and no funding — proof that a serious on-device research model can be made in your own lab. If you build small, honest, and on-device, you don't need a data-center to do meaningful work.