Cooked by Swarm and Bee LLC on sovereign Blackwell compute, for the OpenDiabetic mission: give people what they need, not what you have. Free, open, defendable.

Beat-base — proven, not asserted

Held-out perplexity on diabetic/medical text the model was never trained on (the only honest test of domain learning):

	held-out loss	perplexity
Base Qwen3.6-27B	1.6738	5.333
DiabeticDaily-27B	0.7197	2.054
Δ	−0.954 (+57% better)

Verdict: BEAT BASE ✅. A definitive domain-absorption signal — the model models diabetic/medical language 57% better than its base. (Receipt: beat_base_result.json.)

How it was cooked

Base: Qwen/Qwen3.6-27B (Apache-2.0).
Data: the OpenDiabetic deeded corpus — clinician-grade diabetic & medical instruction data (publicly donated at diabeticdatasets.com, PII-scrubbed + verified).
Recipe (gold standard): LoRA r32/α16 on attn+mlp (not the linear-attention state-mixers), LR 1e-5, cosine, early-stopping as the overcook guard. Merged to bf16.

The ladder

🐝 HIVE   DiabeticDaily-27B  +57%   ← you are here (the foundation)
🏠 HOME   DiabeticDaily-9B       +40.7%
🛏️ EDGE   DiabeticDaily-4B     +40.4%  (runs on a $249 Jetson, on-box)

Use it

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("SwarmandBee/DiabeticDaily-27B", trust_remote_code=True)
m = AutoModelForCausalLM.from_pretrained("SwarmandBee/DiabeticDaily-27B", torch_dtype="bfloat16",
        attn_implementation="sdpa", trust_remote_code=True)

Or serve fast via ollama (Q4_K_M GGUF) — see the -GGUF companion repo.

⚠️ Not medical advice

DiabeticAnchor is a diabetic lifestyle, education, and organization model. It does not diagnose, prescribe, or replace a care team. For emergencies, call 911. Educational use only.