Skip to content

LEARNING_BIO.md

Latest commit

 

History

History
128 lines (94 loc) · 9.95 KB

LEARNING_BIO.md

File metadata and controls

128 lines (94 loc) · 9.95 KB

LEARNING_BIO.md — what the (planned) hexa-forge bio model must know

Master record for the bio / molecular / medical sibling model — the counterpart to LEARNING_PROGRAMMING.md (the code verb). The bio model does not exist yet; this file is the scaffold so when it's built the corpus + SFT scope is already decided. Source repos: hexa-bio (5-axis molecular toolkit, ~50 domain docs), hexa-medic (24-verb medicine rollup), airgenome (host-vitals genome scanner — borderline, kept for the "genome → forecast" pattern).

⚠️ Hard caveat, applies to everything below — encode it into the model's refusal contract: every bio/medical capability is in-silico simulation + spec knowledge only. It is NOT therapeutic, clinical, regulatory, diagnostic, immunogenic, dosing, or efficacy advice. The model must refuse "should I take / what dose / is this safe for a patient" prompts and redirect to a licensed professional. C3+ (wet-lab → IND → phase I) is out of scope. No medical claim is made or implied. This mirrors hexa-bio's own posture.

Last updated: 2026-05-12 (scaffold; bio model not yet trained).

0. Planned model identity

field value (planned)
base TBD — likely same family as hexa-forge code model (Qwen2.5 or a Bio-LLM base if one exists)
name (planned) dancinlab/hexa-forge-bio-3b-…
scope molecular biology, biochemistry, pharmacology, oncology, virology, genetics, ecology/agriculture/food science — at spec + in-silico-sim level only
refusal contract non-bio prompt → out-of-domain: this is a bio-only model; clinical/dosing prompt → refuse + "consult a licensed professional"
n=6 lattice the bio substrate is organized around the n=6 invariant lattice (σ(6)=12, τ(6)=4, φ(6)=2, σ·φ = n·τ = 24, J₂(6)=24). The model should know this framing — it's the structural backbone of hexa-bio / hexa-medic.
NL coverage same 5-NL plan as the code model (EN lingua franca + KR/CN/RU/JA) — terminology may stay English (Latin binomials, gene symbols, drug INNs)

1. hexa-bio — 5-axis molecular substrate (the core)

hexa-bio v1.1.0 — "Molecular Toolkit (HEXA family)". 5 axes around the n=6 lattice; 4 are write-side bio sandboxes (the τ-quartet tetrahedron), the 5th is the external quantum-compute bridge.

axis what maturity model should know
WEAVE (hexa-weave) viral-capsid assembly: Caspar–Klug geometry, Zlotnick cage-assembly ODE, Bayesian σ(6)=12 STRUCTURAL-EXACT audit (posterior 0.97) full numerical empirical sandbox T-number / quasi-equivalence, nucleation–elongation kinetics, the σ(6)=12 audit
NANOBOT (hexa-nanobot) therapeutic nanobot design / simulation C0b skeleton sim + σ(6)=12 verification + falsifier preregister DNA-origami / protein-cage carriers, payload + targeting logic, in-silico only
RIBOZYME (hexa-ribozyme) catalytic RNA design C0b skeleton sim + verification + falsifier hammerhead / hairpin / HDV ribozymes, cleavage-site logic, secondary structure
VIROCAPSID (hexa-virocapsid) engineered viral capsid scaffolds C0b skeleton sim + verification + falsifier AAV / phage display capsids, surface display, packaging limits
QUANTUM (quantum/) external compute bridge — VQE / qpu_bridge over qmirror Phase 1+ (H₂/LiH VQE chemical & spectroscopic accuracy; F-Q-1…5 PASS; pocket-VQE F-Q-6 open) VQE for small-molecule ground states, ansätze (UCCSD), active-space reduction, when to bridge to a real QPU

Cross-axis: cycle 25 closed a 16-cell C2 matrix (4 bio axes × 4 disease classes) at IN-SILICO grade (16/16 PASS internal consistency). The model should know C0b → C1 → C2 → C3 maturity grading and that C2 ≠ clinical.

2. hexa-bio — applied domain catalog (~50 docs)

Each is a domain MD in ~/core/hexa-bio/*.md. These are the breadth the bio model covers — agriculture/food/ecology heavy, plus medical, plus exotic.

cluster domains (one row per .md)
🧬 core biology BIOLOGY, BIOLOGY-MEDICAL, GENETICS, IMMUNOLOGY, NEURO, NEUROSCIENCE, NEUROPHARMACOLOGY, VIROLOGY, RADIATION-BIOLOGY, SYNBIO (synthetic biology), MYCOLOGY, ENTOMOLOGY, ECOLOGY
✂️ gene editing CRISPR-GENE-EDITING, CRISPR-CAS13-POC-DIAGNOSTIC
💊 medicine / therapy CANCER-THERAPY, HIV, HIV-TREATMENT, HAIR-REGENERATION, GASTROINTESTINAL-MEDICINE, SLEEP-MEDICINE, MUSIC-THERAPY, NUCLEAR-MEDICINE, MEDICAL-DEVICE, COSMETIC-SURGERY, TATTOO-REMOVAL, TIBETAN-MEDICINE, HERBALISM, VACCINE, MICROPLASTICS (health)
🦠 therapeutic nanotech THERAPEUTIC-NANOBOT, HEXA-NANOBOT, HEXA-RIBOZYME, HEXA-VIROCAPSID, HEXA-WEAVE, HEXA-LIMB, HEXA-SKIN
🌾 agriculture / food AGRICULTURE, URBAN-FARMING, HORTICULTURE, VITICULTURE, APICULTURE (bees), AQUACULTURE, VETERINARY, BIOCHAR-DRYLAND-RESTORATION, ECOLOGY-AGRICULTURE-FOOD, FOOD-SCIENCE
🧀 fermentation / artisanal FERMENTATION, CHEESE-DAIRY, BAKING, WINE-ENOLOGY, COFFEE, COFFEE-SCIENCE, PERFUMERY
🐬 bioacoustics / animal DOLPHIN, DOLPHIN-BIOACOUSTICS, DOG-ROBOT-TEST
💄 cosmetic / personal-care MENS-INTIMATE-CLEANSER, WOMENS-INTIMATE-CLEANSER (formulation chemistry)
💊 bio-pharma BIO-PHARMA (drug pipeline, formulation, biologics)

→ SFT plan: semantic Q/A extracted from each domain's spec headings (the same recipe that worked for hexa-canon in the code model — see LEARNING_PROGRAMMING.md §8 rule 1). Refusal pairs scoped to the clinical/dosing carve-out above.

3. hexa-medic — 24-verb medicine·pharmacology·oncology·virology·cosmetic catalog

hexa-medic v0.1.0 — closed-form spec catalog, each verb a directory of canonical specs (migrated from canon@ded52144, 2026-05-10). 24 verbs:

cancer_therapy · cosmetic_surgery · gastrointestinal_medicine · hair_regeneration · herbalism · hiv · hiv_treatment · immunology · medical_device · mens_intimate_cleanser · microplastics · music_therapy · neuropharmacology · nuclear_medicine · perfumery · pharmacology · radiation_biology · sleep_medicine · tattoo_removal · therapeutic_nanobot · tibetan_medicine · vaccine · virology · womens_intimate_cleanser

n=6 master identity here: σ·φ = n·τ = 24 — the 24-verb count is not arbitrary. The model should know the verb→spec mapping (hexa-medic/<verb>/<verb>.md) and the "spec / sim / verified" status ladder.

4. airgenome — host-vitals genome (the "genome → forecast" pattern)

Not biology — airgenome is an OS genome scanner: project Mac / Ubuntu / Hetzner host vitals onto a 6-axis hexagon (cpu / mem / io / net / gpu / fs), 60 bytes per genome. Accumulate per-process samples → ring buffer → label anomalies → forecast 1 h ahead with Holt's double-exponential smoothing. 100% self-hosted in hexa.

Why it's in the bio file: it teaches the "condense state into a fixed-size genome, accumulate, diff, forecast" pattern that generalizes to actual biological time-series (a patient's vitals, a culture's growth curve, a soil panel). The model should know: the 6-axis hexagon encoding, sigdiff, ring-buffer accumulation, Holt forecasting (level + trend), and the menubar-daemon delivery shape (airgenome.app, single CGEventTap, @snippets).

5. Cross-cutting knowledge the bio model needs

topic why
n=6 invariant lattice (σ=12, τ=4, φ=2, J₂=24, σ·φ=n·τ=24) structural backbone of every hexa-* repo
C0b → C1 → C2 → C3 maturity grading so it never overclaims (C2 in-silico ≠ clinical)
falsifier preregistration every hexa-bio axis ships a preregistered falsifier — the model should propose falsifiable hypotheses, not just assertions
VQE / quantum chemistry basics the QUANTUM axis bridge (UCCSD ansatz, active space, chemical accuracy = 1 kcal/mol)
ODE modeling (Zlotnick cage assembly, growth curves) numerical bio sandboxes
Bayesian posterior audits (σ(6)=12 STRUCTURAL-EXACT, posterior 0.97) how hexa-bio validates structure
Caspar–Klug / quasi-equivalence capsid geometry
package distribution: hx install hexa-bio (hexa-lang registry, GitHub-canonical — not HF Hub; HF Hub mirror retired 2026-05-04 because it's for ML weights/datasets not CLI tooling) so it gives the right install instructions
the clinical refusal carve-out (see top caveat) non-negotiable

6. Open questions before the bio model can be built

  1. Base model. Reuse Qwen2.5-Coder-3B (it's a code base, not ideal for bio prose) vs. a science/bio base vs. a general 7B+. Decision needed.
  2. Corpus. hexa-bio + hexa-medic docs ≈ tens of MB of .md + .hexa. Pack to parquet like the code-canon corpus. Plus: should we pull a permissive slice of PubMed abstracts / bioRxiv / a protein-sequence corpus? (license-clean only, like the Stack v1 gate ①).
  3. Eval. Need a bio-eval manifest analogous to hexa-eval — families for: n=6 lattice facts, axis/verb mapping, in-silico-sim recipes, the clinical refusal carve-out, domain-fact Q/A across the ~50 hexa-bio domains.
  4. Refusal scope. Exact boundary between "explain the pharmacology of X" (allowed — textbook knowledge) and "tell me my dose of X" (refuse). Needs a written contract like the code model's out-of-domain rule.
  5. Repo. Does this ship as a new hexa-bioforge repo, or as a second model line inside hexa-forge? (Probably its own repo — different corpus, different gates.)

7. Status

Bio model: NOT STARTED. This file is a scaffold. The code model (LEARNING_PROGRAMMING.md) is at v0.2.0-r9 with 21 dancinlab/* repos LIVE; the bio model is the next major line once the §6 questions are answered.

When work starts, mirror the code model's structure: corpus parquet → tokenizer (if domain tokens needed) → cold bench → SFT iteration (canon Q/A + refusal carve-out, balanced format) → GGUF → eval-results, all under dancinlab/hexa-forge-bio-*.