BeeNara v0.4: 332MB local doc sorter hits 91.5% on unseen folders
razer_psycho · reddit · 2026-10-28
BeeNara v0.4 (Apache-2.0) is a 332MB CPU model that files documents locally or says 'none fits'. Accuracy on 1,500 unseen-folder tasks rose 70% → 91.5% across versions. A Learn-then-Test threshold certifies ≤1% misfiling at 90% confidence; v0.4 auto-files 59% of tasks with only 0.3% wrong. A new Qwen-2B 'Big' sibling reads whole folder trees (600 folders) plus the page image, needing 4.3GB GPU in 8-bit. Key lesson: label smoothing destroyed the 'ask when unsure' behavior. All training data is synthetic, generated by two local teacher models cross-checking each other across 32 document types with real OCR noise.
More from coding & agent
- Petabyte pitches sleep-when-idle hosting for 24/7 AI agents with event-triggered wakeups — Petabyte-Cloud · 2026-10-28
- Claude built its own Obsidian plugin, and its human is reviewing AI-edited videos on iPhone — evielync · 2026-10-28
- Dev stitches Claude, GPT, Grok and Gemini into one multi-agent computer-use system — RileyRalmuto · 2026-10-28
- One illustration to a full VTuber model: dev builds open-source kit using Claude and Codex agents — bdsqlsz · 2026-10-28
- Ahrefs Podcast: How Ryan Law Vibe-Coded a Full Website in One Day — 5le · 2026-10-28
- Dev explains Microsoft's four Copilots as nested if, for, while and while(true) — pswider · 2026-10-28