BeeNara: a 332MB CPU-only classifier that knows when no folder fits (96.8% recall)
razer_psycho · reddit · 2026-09-26
Reddit user razerpsycho released BeeNara, a tiny 332MB ONNX cross-encoder for local document sorting.
- Motivation: small local LLMs like Qwen3.5-4B tend to hallucinate a category instead of admitting "none fits", while large models are overkill for basic sorting.
- How it works: zero-shot classification against plain-text folder names, with split-conformal prediction for calibrated confidence — uncertain or unmatched documents return "none fits" for human review.
- Performance: runs fully offline on a laptop CPU in 0.2–0.3s per document (no PyTorch/GPU), bilingual English/German, and catches 96.8% of documents whose correct folder is missing.
Model and code are on Hugging Face (Kwokou/BeeNara), usable standalone in Python.
More from Research
- Silent mistakes that can hurt your Numerai models, per new community writeup — matlabulous · 2026-09-26
- Zenke Lab to present spiking network and surrogate gradient work at Bernstein Conference — hisspikeness · 2026-09-26
- Michael Levin: intelligence doesn't require genes, cognition is widespread in life — drmichaellevin · 2026-09-26
- Position paper argues grounded abduction is the missing reasoning leap for LLMs — bibryam · 2026-09-26
- HKUST benchmark: even best legal AI agents hallucinate in 89% of trajectories — 机器之心 · 2026-09-26
- InternLM open-sources Intern-Decision 4B/0.8B: structured decisions in one forward pass — jacek2023 · 2026-09-26