Beat-base — proven
Held-out perplexity vs base Qwen3.5-4B (text never trained on):
Table with columns: held-out loss, perplexity | held-out loss | perplexity |
|---|
| Base Qwen3.5-4B | 1.5062 | 4.510 |
| DiabeticBaby-4B | 0.8982 | 2.455 |
| Δ | −0.608 (+40.4% better) | |
Verdict: BEAT BASE ✅. A 4B that models diabetic/medical language ~40% better than base — and at Q4 it's
~2.6GB, running at usable speed on a Jetson with PHI never leaving the box.
How it was cooked
- Base: Qwen/Qwen3.5-4B (Apache-2.0). Data: the same deeded OpenDiabetic corpus as the 27B/9B.
- Recipe: LoRA r32/α16 on attn+mlp, LR 2e-5 (small-model tier), 0.7ep, early-stop. Merged bf16.
Run it on a Jetson (Q4 GGUF, ollama) — see the -GGUF companion repo
ollama create diabetic-baby -f Modelfile # FROM diabeticbaby-4b-q4_k_m.gguf
ollama run diabetic-baby "What's a good diabetic breakfast?"
This is the brain behind the LocalDiabetic edge node — sovereign, private, free.
The ladder: 🐝 27B anchor (+57%) → 🏠 9B home (+40.7%) → 🛏️ 4B edge (+40.4%)
⚠️ Not medical advice — diabetic lifestyle/education/organization only. Not a diagnosis. Emergencies → 911.
© 2026 Swarm and Bee LLC · opendiabetic.com · Apache-2.0 · We slow cook the truth. 🐝