Cookbook · Diabetes Companion · 5000
**Target failure mode.** Full-stack diabetes assistant. Clinical decision support + research grounding + patient communication + lived-experience register, all in one cooked adapter. **Two-stream architecture preserved** — the cooked model never uses lived voice as medical authority, never uses sourced authority as register.
This is the master cookbook. **5000 cells is the honest "confident domain pivot" floor** — at this scale the cooked model is reliably a diabetes specialist, not just a base model with a glycemic-shaped patch.
This is the recipe staged for the dmack.ai senior-hack cook itself. **If you order this, you are ordering exactly what we are cooking for ourselves.**
The receipt
**This recipe is the dmack.ai cook.**
The dmack.ai project is the Swarm & Bee flagship medical model — Donovan's lived experience as a Type 1 diabetic + Addison's disease, fused with peer-reviewed medical authority, in one model. The ingredient mix below is **the exact 5000-cell mix sealed for the senior-hack canary-then-cook review.**
**Status:** staged. The full cook is in the canary-then-cook queue. When the cook finalizes, the actual loss curve, full 60-probe eval delta, Hedera-anchored bundle hash, and the cooked adapter weights publish in this cookbook.
**Estimated completion:** 2026-Q3.
**Until then:** order this only if you are willing to be the first external receipt. You'll get the corpus identical to what we are cooking on, you can cook it yourself, and your eval-delta becomes part of the cookbook's published history.
The recipe — Swarm & Bee Gold Standard QLoRA
Same Gold Standard config used to cook Atlas-Qwen-27B (loss 0.4186) and SwarmCurator-9B (loss 0.707). See [Glycemic Reasoning](/cookbooks/glycemic-reasoning.md) for full YAML.
**Master-cookbook adjustments:**
epochs: 2 # was 3 — more data, less overfit risk; 2 passes is plenty at 5000 cells
gradient_checkpointing: true # mandatory at this size on most consumer 5090/PRO 6000 setups
For 4B base: ~10-14h cook on a single RTX PRO 6000 Blackwell.
For 8B base: ~20-28h on the same.
For 27B base: cook on a 2× PRO 6000 with FSDP — ~36-50h.
The ingredients — 5000 cells, deterministic
Cell selection is **deterministic per cookbook** (seed = sha256("diabetes-companion"), sha256-pinned). The 2,500 Dmack's Royal Jelly cells in this cookbook are sampled across all 10 source streams in the master, in their original proportions — you get a representative slice of the full multi-source flagship, not a biased pull.
Schema · multi-source
Cells in this cookbook carry their original metadata (specialty, source, tier, citation if applicable). The shared minimum is `{question, answer, specialty, tier, source}`. The full schema-by-source is documented in the per-SKU cookbooks ([Glycemic Reasoning](/cookbooks/glycemic-reasoning.md), [Diabetic Foot Care](/cookbooks/diabetic-foot-care.md), [Patient Communication](/cookbooks/patient-communication.md)).
**Two-stream tag:** each cell carries a `source` field. Cells from `dmack-hand-curated` (Founder's Pretzel) are tagged as such. If you want to *exclude* the lived-voice stream at training time (e.g. for a pure-authority research model), filter `metadata.source != "dmack-hand-curated"` in the loader. The loss-weighting trick for over-emphasizing the lived stream also works — your choice.
Loader — multi-source-aware
import json
def load_cookbook(jsonl_path: str, exclude_sources: list[str] = None) -> list[dict]:
"""Multi-source-aware loader for the Diabetes Companion 5000 bundle.
Pass exclude_sources to filter out a stream (e.g. for pure-authority training)."""
exclude = set(exclude_sources or [])
cells = []
with open(jsonl_path) as f:
for line in f:
o = json.loads(line)
if o.get("source") in exclude:
continue
cells.append({
"instruction": o["question"],
"response": o["answer"],
"stream": "lived" if o.get("source") == "dmack-hand-curated" else "sourced",
"metadata": {k: v for k, v in o.items() if k not in ("question","answer")},
})
return cells
The `stream` field is useful for per-sample weighting if you want to bias the cook toward one stream or the other. Default config treats both streams equally.
Eval · dmack_eval_set_v1 (all 60 probes)
This is the only cookbook that runs the **full 60-probe eval set**. Every category in scope:
This eval set was designed against this exact recipe. Running it post-cook tells you if the model became dmack.ai or if it became something else.
How to order
swarmbee-bakery cookbook diabetes-companion
swarmbee-bakery order \
--sku cookbook \
--cookbook diabetes-companion \
--name "Your Name" \
--email "[email protected]" \
--settlement swarmusdc \
--notes "cooking on Qwen-3.5-8B · I'm prepared to be the first external receipt and share my eval delta back" \
--confirm
If you're willing to publish your pre/post eval delta back to us, we **waive the cooked-for-you fee entirely**. Skin-in-the-game pricing for first customers of a pre-publication recipe.