Large Botanical Model Proposed architecture · nothing built Methods peer-reviewed Calculators illustrative

Public botanical data are too small to train a model, so the proposal adapts open ones

“Large Botanical Model” names a proposed architecture: open models adapted to EUTECNIQ's own paired lot records. EUTECNIQ has no model, no training data and no compute contract.

738,794natural-product structures in the COCONUT October 2026 release

48.2 millionSMILES tokens in all of them, by our count

≈29,000×that count trained Chinchilla 70B, a general language model (1.4 trillion tokens)

38 of 59botanical markers meet both the Lipinski and Veber limits (PubChem computed)

Status. A proposal: no laboratory, LIMS records or trained model exist. Public counts and constants were checked on October 22–24, 2026.

Academy fundamentals: Course 11, Multi-objective ingredient design

Architecture Proposal

Nine data types would meet in one shared embedding, trained so that records of the same lot sit together

Proposed Large Botanical Model: data types, encoders, shared embedding, tasks and the wet-lab loopProposed architecture

Scroll sideways inside the diagram to see all four columns.

Select a component in the diagram (click, or Tab and press Enter) to read what it would do, where it starts from and its status. The same information is in the text table below the diagram.

Component basis: retrieval-augmented generation (Lewis et al. 2020); LoRA (Hu et al. 2021); OLMo 2 (Team OLMo 2024, Apache 2.0 model card); MoLFormer (Ross et al. 2022); ChemBERTa (Chithrananda et al. 2020); DreaMS (Bushuiev et al. 2025); CReSS (Yang et al. 2021); plant barcodes (CBOL Plant Working Group 2009; Chen et al. 2010); contrastive alignment (van den Oord et al. 2018; Radford et al. 2021). Whether these methods transfer to botanical lots is the hypothesis the roadmap tests. Dashed outline: EUTECNIQ's own data and laboratory, which do not exist yet.

All 23 components as text: role, starting point and status

Components of the proposed Large Botanical Model
Component	What it would do	Starting point or source (verified)	Status
Literature text	Supplies evidence passages for answers and the links between species, constituents and studies.	PubMed lists more than 40 million citations and abstracts (NLM). Full texts only where the license permits; PMC states that license terms vary by article and that bulk retrieval is allowed only through its four automated services.	Public source
Monographs	Supplies identity tests, acceptance criteria and botanical descriptions for retrieval.	Compendial texts are published under copyright by their publishers; use in an index or for training would be licensed first. Terms have not been assessed.	License needed
Regulatory documents	Supplies the regulation and guidance text that answers are checked against.	Works of the U.S. Government are not protected by copyright under 17 U.S.C. §105, which covers the Code of Federal Regulations and FDA guidance documents.	Public source
Chemical structures	Supplies natural-product chemical space and the species in which each structure was reported.	COCONUT, October 2026 release: 738,794 molecules, CC0. LOTUS: 750,000+ referenced structure–organism pairs on Wikidata, CC0.	Public source
MS/MS spectra	Supplies fragment spectra for structure retrieval and for comparing lots.	MassBank: 119,845 spectra of 18,529 compounds; each record carries a mandatory Creative Commons license field (CC0, CC BY, CC BY-SA, CC BY-NC or CC BY-NC-SA). GNPS/MassIVE public datasets: terms not reviewed here.	Public source
NMR spectra	Confirms structures and profiles whole extracts.	Mostly EUTECNIQ's own spectra of authenticated reference lots; no public NMR collection was assessed for this page.	To be generated
Chromatograms	Fingerprints of lots for identity and consistency.	EUTECNIQ's own HPLC and HPTLC runs under recorded methods; public chromatogram collections were not assessed.	To be generated
DNA barcodes	Species-level identity of the source plant, where intact DNA survives processing.	rbcL + matK is the recommended core plant barcode (CBOL Plant Working Group 2009); ITS2 was validated for medicinal plant species (Chen et al. 2010).	Public method
EUTECNIQ LIMS records	Supplies the pairs that make alignment possible: spectra, chromatograms, barcodes and assay results of the same authenticated lot.	None. EUTECNIQ has no laboratory or LIMS yet; these records would be created from the first sample, with provenance.	Does not exist yet
Language model and retrieval index	Drafts answers from retrieved passages; adapted to botanical vocabulary with low-rank adapters rather than pretrained from scratch.	OLMo 2 7B: Apache 2.0, trained on 4 trillion tokens, with weights, training data and code released (model card). LoRA (Hu et al. 2021); retrieval-augmented generation (Lewis et al. 2020).	Candidate
Chemical language encoder	Turns a SMILES string or a molecular graph into a vector.	MoLFormer was pretrained on SMILES of 1.1 billion molecules from PubChem and ZINC (Ross et al. 2022). ChemBERTa released a curated set of 77 million PubChem SMILES (Chithrananda et al. 2020).	Candidate
Spectrum transformer	Turns an MS/MS spectrum into a vector that can be compared with structures.	DreaMS was pretrained without labels on millions of tandem mass spectra (the GeMS dataset, mined from MassIVE GNPS) by predicting masked peaks and retention orders; the authors released the dataset and model (Bushuiev et al. 2025).	Candidate
NMR encoder	Places ¹³C NMR spectra and structures in one space for cross-modal retrieval.	CReSS used deep contrastive learning to retrieve structures from ¹³C NMR spectra (Yang et al. 2021, Anal Chem).	Method published
1-D signal encoder	Encodes aligned chromatographic traces.	No pretrained starting point is assumed; it would be trained on EUTECNIQ's own runs.	To be built
Sequence encoder	Encodes barcode sequences for species-level matching.	k-mer frequency profiles are a simple first representation; reference sequences from public libraries, after their terms are confirmed.	To be built
Lot-record encoder	Encodes assay, impurity, stability and process results of a lot.	EUTECNIQ data only, under 21 CFR Part 11 controls where the record is required.	To be built
Retriever and citation verifier	Selects passages, conditions the language model on them, and removes any sentence its cited passage does not support.	Retrieval-augmented generation (Lewis et al. 2020); citation-quality metrics of the ALCE benchmark (Gao et al. 2023).	Candidate
Shared embedding	One vector space in which paired records of the same lot are close; an applicability check refuses inputs far from training data.	Symmetric contrastive loss (InfoNCE, van den Oord et al. 2018; the paired form used by CLIP, Radford et al. 2021).	Proposed
Cited literature answers	Drafts dossier sections and answers scientific questions with a source for every sentence, for expert editing.	Measured by citation precision and recall (Gao et al. 2023). Never filed or published without human review.	Proposed
Identity verification	Ranks how consistent a lot's spectra, fingerprint and barcode are with authenticated references of the labeled species.	A screen only. Part 111 requires at least one appropriate test or examination to verify the identity of each dietary ingredient (21 CFR 111.75(a)(1)), using scientifically valid methods (111.75(h)(1)).	Proposed
Adulterant flagging	Flags lots that resemble known adulterated or substituted material.	A flag triggers a confirmatory test by a validated method; it never releases or rejects a lot.	Proposed
Property prediction	Predicts physicochemical properties with a prediction interval, or abstains outside the applicability domain.	Baselines and calibration rules are on Computational science and AI.	Proposed
Wet-lab validation	Tests every consequential output with a validated method and returns the result to LIMS as a new paired record.	Proposed EUTECNIQ laboratory or a qualified external partner; none is in operation.	Does not exist yet

Explanation, details and sources

Status. A proposal from a company in formation with no laboratory, LIMS records or trained model. Public counts, published constants and hardware figures were checked against primary sources on October 22–24, 2026. The COCONUT token count is our own computation; other calculator numbers illustrate published equations.

A proposed architecture≠an existing model
A scaling-law projection≠a measured result on botanical data
Lower training loss≠better identity, safety or quality decisions
A model flag≠a validated test result

Governance, data rights and funding gates: Computational science and AI. The first narrow model, for extraction yield: Botanical Intelligence.

An encoder turns one kind of record, such as a mass spectrum, into a vector. Text feeds a language model through a retrieval index, so answers come from documents, not memory; measurements feed a shared embedding for three screening tasks.

Symmetric contrastive loss over a batch of B paired recordsℒ = −12B Σ_i=1…B [ ln e^s_ii/τΣ_j e^s_ij/τ + ln e^s_ii/τΣ_j e^s_ji/τ ]
s_ij = u_i · v_j‖u_i‖ ‖v_j‖

u_i is one record's vector (the MS/MS spectrum of lot i) and v_i its partner's (the same lot's chromatogram or structure). The loss is small when each record is most similar to its own partner among B candidates; temperature τ sets how sharply.

InfoNCE: van den Oord, Li, Vinyals 2018, arXiv:1807.03748. Symmetric paired form: Radford et al. 2021, arXiv:2103.00020 (CLIP, 400 million image–text pairs). For botanicals the pairs exist only where one authenticated lot is measured several ways, which is why EUTECNIQ's own records come first.

Tasks Proposal

Each task would answer one question for a named decision owner, and none of them would release, reject or classify a product

From input to model output to a named person and a validated next stepProposed tasks

EUTECNIQ proposal. Quality decisions stay with the quality unit and validated methods (Computational science and AI).

Show the table (4 rows)

The model narrows where scientists look. Quality decisions remain with the quality unit and validated methods.

Task	Input	Output	Decision owner	What happens next
Identity verification	MS/MS, NMR, chromatogram and barcode records of an incoming lot	Similarity to authenticated references of the labeled species, with a calibrated score or an abstention	Qualified analyst	The validated identity test is run as required; a low score adds an orthogonal test
Adulterant flagging	The same records, plus reference records of known adulterants	A flag with the nearest matching adulterant class	Quality unit	Confirmatory test by a validated method; the outcome becomes a labeled training record
Property prediction	Structure, or the measured profile of a lot	Predicted value with an interval, or no prediction outside the applicability domain	Formulation scientist	Candidates are ranked for measurement; predictions never replace data
Cited literature answers	A scientific or regulatory question	A draft in which every sentence links to the passage that supports it	Responsible scientist or regulatory specialist	Expert review and editing; unsupported sentences are removed and logged

Multi-objective selection PubChem data

Screening on several property limits at once shows which botanical markers fall outside drug-derived heuristics

59 botanical marker compounds on six descriptors: drag along any axis to keep only the candidates inside a rangePubChem computed values

PubChem PUG-REST computed properties (Cactvs, XLogP3), retrieved Oct 24, 2026; Kim et al., Nucleic Acids Res 2025;53(D1):D1516–D1525. Violations counted against Lipinski et al. 2001 (Adv Drug Deliv Rev 46:3–26), with XLogP3 standing in for CLogP. Computed descriptors are model estimates, not measurements.

Data table: all 59 compounds

PubChem computed descriptors of 59 botanical marker compounds
Compound	PubChem CID	MW (g/mol)	XLogP3	TPSA (Å²)	H-bond donors	H-bond acceptors	Rotatable bonds	Rule of five and Veber met
Quercetin	5280343	302.23	1.5	127	5	7	1	yes
Curcumin	969516	368.4	3.2	93.1	2	6	8	yes
Berberine	2353	336.4	3.6	40.8	0	4	2	yes
Resveratrol	445154	228.24	3.1	60.7	3	3	2	yes
Caffeine	2519	194.19	-0.1	58.4	0	3	0	yes
Capsaicin	1548943	305.4	3.6	58.6	2	3	9	yes
Piperine	638024	285.34	3.5	38.8	0	3	3	yes
Epigallocatechin gallate	65064	458.4	1.2	197	8	11	4	no
Rosmarinic acid	5281792	360.3	2.4	145	5	8	7	no
Chlorogenic acid	1794427	354.31	-0.4	165	6	9	5	no
Glycyrrhizin	14982	822.9	3.7	267	8	16	7	no
Artemisinin	68827	282.33	2.8	54	0	5	0	yes
Andrographolide	5318517	350.4	2.2	87	3	5	3	yes
Silibinin (silybin)	31553	482.4	2.4	155	5	10	4	no
Hypericin	3663	504.4	5.7	156	6	8	0	no
Hyperforin	441298	536.8	9.6	71.4	1	4	11	no
Withaferin A	265237	470.6	3.8	96.4	2	6	3	yes
Ginsenoside Rb1	9898279	1109.3	0.3	377	15	23	16	no
Baicalin	64982	446.4	1.1	183	6	11	4	no
Ursolic acid	64945	456.7	7.3	57.5	2	3	1	yes
Kavain	5281565	230.26	2.5	35.5	0	3	3	yes
Valerenic acid	6440940	234.33	3.2	37.3	1	2	2	yes
Harpagoside	5281542	494.5	-0.6	175	6	11	7	no
Salidroside	159278	300.3	-0.6	120	5	7	5	yes
Rosavin	9823887	428.4	-2	158	6	10	7	no
6-Gingerol	442793	294.4	2.5	66.8	2	4	10	yes
Apigenin	5280443	270.24	1.7	87	3	5	1	yes
Luteolin	5280445	286.24	1.4	107	4	6	1	yes
Kaempferol	5280863	286.24	1.9	107	4	6	1	yes
Naringenin	439246	272.25	2.4	87	3	5	1	yes
Hesperidin	10621	610.6	-1.1	234	8	15	7	no
Rutin	5280805	610.5	-1.3	266	10	16	6	no
Ellagic acid	5281855	302.19	1.1	134	4	8	0	yes
Gallic acid	370	170.12	0.7	98	4	5	1	yes
Oleuropein	5281544	540.5	-0.4	202	6	13	11	no
Hydroxytyrosol	82755	154.16	-0.7	60.7	3	3	2	yes
Thymol	6989	150.22	3.3	20.2	1	1	1	yes
Menthol	1254	156.26	3	20.2	1	1	1	yes
Eugenol	3314	164.2	2	29.5	1	2	3	yes
Lutein	5281243	568.9	11	40.5	2	2	10	no
β-Carotene	5280489	536.9	13.5	0	0	0	10	no
Forskolin	47936	410.5	1	113	3	7	3	yes
Huperzine A	854026	242.32	0	55.1	2	2	0	yes
Tanshinone IIA	164676	294.3	4.3	47.3	0	3	0	yes
Emodin	3220	270.24	2.7	94.8	3	5	0	yes
Salicin	439503	286.28	-1.2	120	5	7	4	yes
Parthenolide	6473881	248.32	2.3	38.8	0	3	0	yes
Ginkgolide B	11973122	424.4	-0.4	149	3	10	1	no
Bilobalide	73581	326.3	-0.3	119	2	8	1	yes
Honokiol	72303	266.3	5	40.5	2	2	5	yes
Icariin	5318997	676.7	1.7	234	8	15	9	no
Puerarin	5281807	416.4	0	157	6	9	3	no
Genistein	5280961	270.24	2.7	87	3	5	1	yes
Matrine	91466	248.36	1.6	23.6	0	2	0	yes
β-Boswellic acid	168928	456.7	8.3	57.5	2	3	1	yes
Aloin	9866696	418.4	-0.1	168	7	9	3	no
Withanolide A	11294368	470.6	3.3	96.4	2	6	2	yes
Crocin	5281233	977	-2.5	391	14	24	20	no
Allicin	65036	162.3	1.3	61.6	0	3	5	yes

38 of 59 markers meet both heuristics; polar surface area excludes the mostComputed from PubChem data

Counts computed by this page from the table above. Lipinski CA et al. Adv Drug Deliv Rev 2001;46:3–26; Veber DF et al. J Med Chem 2002;45:2615–2623.

Data table: markers meeting each limit

Number of the 59 markers meeting each limit
Limit	Markers meeting it
Molecular weight ≤ 500	48
XLogP3 ≤ 5	53
H-bond donors ≤ 5	44
H-bond acceptors ≤ 10	49
Rule of five, ≤ 1 violation	45
Rotatable bonds ≤ 10	55
TPSA ≤ 140 Å²	41
Veber, both limits	40
Rule of five and Veber	38

Of the 21 markers outside one or both heuristics, 18 exceed 140 Å² polar surface area: candidates for delivery work, not proof of poor bioavailability.

Skeletal formulas of luteolin, kaempferol, quercetin and rutin. The 2-phenylchromen-4-one (flavone) scaffold common to all four is drawn in teal; they differ in hydroxyl pattern, and rutin carries a rutinose disaccharide on the 3-oxygen, which doubles its molar mass from 302.24 g/mol for quercetin to 610.52 g/mol. — Structures: PubChem isomeric SMILES, CIDs 5280445, 5280863, 5280343, 5280805 (retrieved Oct 25, 2026); RDKit 2026.09 2D depiction; molar masses computed with RDKit; scaffold found by SMARTS match. Generated illustration, EUTECNIQ. Full-size SVG

Explanation, details and sources

Multi-objective selection keeps only candidates that satisfy several limits at once; parallel coordinates show every candidate as one line across all axes, and brushing an axis applies a limit. A property model would supply such values, with intervals, for compounds that lack them. Here the values are PubChem's computed descriptors, themselves model estimates, for 59 well-known marker compounds chosen by this page as examples; PubChem's acceptor and donor counts are not the same calculation as the original papers'.

Lipinski's rule of five states that poor absorption or permeation is more likely above 5 hydrogen-bond donors, 10 acceptors, a molecular weight of 500 or a calculated log P of 5. Veber et al. found, in rat data for over 1,100 drug candidates, that 10 or fewer rotatable bonds and a polar surface area of 140 Å² or less predicted a high probability of good oral bioavailability. Both are heuristics from drug-discovery data sets; they do not measure a botanical's absorption. See Course 11 for selection methods.

Tokenization Published method

A molecule becomes a sequence of chemically meaningful tokens before a language model can read it

Atom-level SMILES regular expression (Schwaller et al. 2018)(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|$|$|\.|=|#|-|\+|\\|\/|:|~|@|\?|>|\*|\$|\%[0-9]{2}|[0-9])

SMILES string (edit or choose an example)

Characters48

Tokens48

Distinct tokens8

Atom tokens22

Bracketed atoms0

Ring-closure digits6

Quercetin (PubChem CID 5280343)Published tokenizer

C1=CC(=C(C=C1C2=C(C(=O)C3=C(C=C(C=C3O2)O)O)O)O)O

Catom=bond or other symbol(branch1ring closure

The two quercetin examples describe the same molecule (identical InChIKey REFJWTPEDVJJIY-UHFFFAOYSA-N) yet give 48 and 42 tokens. A model sees different sequences unless strings are canonicalized or deliberately enumerated.

Regular expression: Schwaller P, Gaudin T, Lányi D, Bekas C, Laino T. Chem Sci. 2018;9(28):6091–6098, Section 3.1; exact string from the authors' smi_tokenizer (github.com/pschwllr/MolecularTransformer). Example SMILES: PubChem compound records, accessed October 22, 2026; the aromatic quercetin form was written by RDKit. Tokens are colored by type: atoms, bonds, branches and ring closures.

Quercetin's SMILES string splits into 48 tokens: 22 atoms, 6 ring-closure digits and 20 bond or branch symbolsReference structure, computed tokens

Structure and SMILES: PubChem CID 5280343 (retrieved Oct 25, 2026). Tokenizer: Schwaller P, et al. Chem Sci 2018 (atom-level SMILES regular expression). Generated illustration, EUTECNIQ. Full-size SVG

Method, assumptions and limits: SMILES tokenizer

What it calculates: The list of tokens produced by the published regular expression, and counts of characters, tokens, distinct tokens, atom tokens, bracketed atoms and ring-closure digits.
Equation: Tokens are the successive matches of the expression shown above, applied left to right with global matching. A string is reproduced exactly when the tokens joined together equal the input, the check the authors' code asserts.
Variables and units: Counts only. Atom tokens are bracketed atoms plus the organic-subset symbols B, C, N, O, P, S, F, Cl, Br, I and their aromatic lower-case forms.
Assumptions: The input is a SMILES string. The tokenizer does not check chemical validity; RDKit or another parser would do that before training.
Valid range: Any text up to 2,000 characters. Characters the expression does not cover are listed in a warning, and the string is then not reproduced exactly.
Does not establish: That a token sequence is the best input for a model; graph representations and other tokenizers exist. Token counts depend on how the SMILES was written, as the two quercetin forms show.
References: Schwaller et al. 2018 (Chem Sci); Schwaller et al. 2019, ACS Cent Sci 5(9):1572–1583 (Molecular Transformer, which reuses the same expression); PubChem.
Verification: An independent Python reference (re module, the same published expression, RDKit for InChIKeys) in tests/numerics/ref/insights-large-botanical-model.py reproduces every count for all seven examples.

The whole COCONUT collection is about 48 million SMILES tokens, a small corpus by language-model standards

All 738,794 COCONUT structures (October 2026): median 58 tokens, 381 distinct tokens.

COCONUT molecules by SMILES length in tokensOur computation

COCONUT CSV lite, October 2026 release (CC0), canonical_smiles column, tokenized with the Schwaller et al. 2018 expression by src/modules/lbm/compute/coconut_tokens.py. Mean 65.3 tokens; 5th to 95th percentile 26 to 129; longest 1,134. No COCONUT records are redistributed here.

Histogram data

Tokens per molecule	Molecules	Share (%)
0–9	1,236	0.2
10–19	13,302	1.8
20–29	38,114	5.2
30–39	77,881	10.5
40–49	121,068	16.4
50–59	134,121	18.2
60–69	107,801	14.6
70–79	73,512	10.0
80–89	52,402	7.1
90–99	31,882	4.3
100–119	38,966	5.3
120–149	28,266	3.8
150–199	15,352	2.1
200–299	4,104	0.6
300–499	693	0.1
500 or more	94	0.0

Training tokens: COCONUT against published language models (log scale)Published figures

Hoffmann et al. 2022, Tables 1 and 3; OLMo 2 7B model card (allenai/OLMo-2-1124-7B); Grattafiori et al. 2024, The Llama 3 Herd of Models, Section 1. SMILES tokens and English text tokens are different units; the comparison shows scale only.

Chart data

Corpus	Tokens	Multiple of COCONUT	Source
COCONUT, all SMILES (our count)	4.82 × 10⁷	1	EUTECNIQ computation on the COCONUT Oct 2026 release
Compute-optimal data for a 1B model	2.02 × 10¹⁰	419	Hoffmann et al. 2022, Table 3 (Approach 1)
Chinchilla 70B training data	1.40 × 10¹²	29,028	Hoffmann et al. 2022, Table 1
OLMo 2 7B pretraining data	4.00 × 10¹²	82,938	OLMo 2 7B model card
Llama 3 405B pretraining data	1.56 × 10¹³	323,460	Llama 3 Herd of Models, Section 1

Explanation, details and sources

SMILES writes a structure as a line of text; a tokenizer cuts it into one token per atom, bond, branch or ring-closure digit, keeping bracketed atoms such as [C@@H] whole. The regular expression in the tokenizer above is Schwaller et al.'s (2018), copied exactly from the authors' code.

Scaling Published model Illustration

Published scaling laws show loss falling with both parameters and data, so a small corpus caps what a larger model can learn

Parametric loss (Hoffmann et al. 2022, Eq. 2 and 10)L(N, D) = E + AN^α + BD^β
E = 1.69, A = 406.4, B = 410.7, α = 0.34, β = 0.28

Compute-optimal frontier (Eq. 4) with C ≈ 6NDN_opt(C) = G (C/6)^a, D_opt(C) = G⁻¹ (C/6)^b
G = (αA/βB)^1/(α+β), a = βα+β, b = αα+β

Fitted constants Compute budget C (FLOPs)1.00 × 10²¹ Unique training tokens available4.82 × 10⁷

Optimal parameters N_opt1.82 × 10⁹

Optimal tokens D_opt9.14 × 10¹⁰

Tokens per parameter50.1

Loss at the optimum2.329

Best loss with the data cap4.611

Data-only floor E + B/D^β4.589

The optimum needs 1,894 times more unique tokens than are available, so this budget is limited by data.

Assumptions

Constants were fitted to one text corpus (MassiveText) with one tokenizer; the loss values do not transfer to SMILES or botanical text
One pass over the data; repeating data is discussed below
C ≈ 6ND counts training FLOPs only

Fitted loss over model size and training tokens, with the compute-optimal frontierReproduced published fit

Drag to rotate, or use the buttons. Amber: models trainable within the budget and the available unique tokens.

Surface: L(N, D) with the constants selected at left (Hoffmann et al. 2022, arXiv:2203.15556, Eq. 10; Besiroglu et al. 2024, arXiv:2404.10102, Table 1). Frontier: Eq. 4. Markers: Chinchilla (70B, 1.4T tokens) and Gopher (280B, 300B tokens) from Hoffmann et al. Table 1, placed at the fitted loss, not at their reported loss. Reproduced; EUTECNIQ has not run this experiment.

Method, assumptions and limits: scaling-law explorer

What it calculates: The compute-optimal model size and token count for a compute budget, the fitted loss there, the best loss reachable when unique tokens are capped, and the loss floor set by the data term alone.
Equations: L(N, D) = E + A/N^α + B/D^β; N_opt = G(C/6)^a, D_opt = G⁻¹(C/6)^b with G, a and b as above (Hoffmann et al. 2022, Eq. 4). With a cap D_avail < D_opt, the one-epoch model uses D = D_avail and N = C/(6D_avail), because loss falls monotonically with N. Floor = E + B/D_avail^β.
Variables and units: N in parameters; D in tokens; C in floating-point operations; L in nats per token on the fitting corpus.
Assumptions: The fits describe dense transformer language models trained on one English-dominant corpus. The two constant sets disagree. Besiroglu et al. refitted the model to data reconstructed from the paper's figures, found the published estimates inconsistent with Hoffmann's own Approaches 1 and 2, and report constants that imply about 20 tokens per parameter, against about 70 for the published Eq. 10. Both are shown so the uncertainty stays visible.
Valid range: The original runs spanned 70 million to over 16 billion parameters and 5 to 500 billion tokens; values far outside are extrapolations. Budgets 10¹⁷–10²⁶ FLOPs; caps 10⁶–10¹⁴ tokens.
Does not establish: Any loss, accuracy or benefit for a botanical model. Repeated data change the picture: Muennighoff et al. (2023) found up to 4 epochs of repetition gave negligible loss changes compared with unique data, while further repetition gave diminishing returns, so a small corpus can be reused only a few times.
References: Hoffmann et al. 2022; Besiroglu et al. 2024; Kaplan et al. 2020 (C ≈ 6N per training token); Muennighoff et al. 2023, arXiv:2305.16264.
Verification: The Python reference finds the optimum by bounded numerical minimization of L(N, C/6N) (scipy) instead of the closed form, and checks readouts and the static table below.

Compute-optimal frontier table for both constant sets

Budget C (FLOPs)	Hoffmann N	Hoffmann D	Hoffmann D/N	Hoffmann loss	Replication N	Replication D	Replication D/N	Replication loss
1.00 × 10¹⁸	8.06 × 10⁷	2.07 × 10⁹	26	3.535	8.05 × 10⁷	2.07 × 10⁹	26	3.490
1.00 × 10¹⁹	2.28 × 10⁸	7.31 × 10⁹	32	2.986	2.62 × 10⁸	6.36 × 10⁹	24	2.927
1.00 × 10²⁰	6.45 × 10⁸	2.58 × 10¹⁰	40	2.600	8.53 × 10⁸	1.95 × 10¹⁰	23	2.553
1.00 × 10²¹	1.82 × 10⁹	9.14 × 10¹⁰	50	2.329	2.78 × 10⁹	6.00 × 10¹⁰	22	2.306
1.00 × 10²²	5.16 × 10⁹	3.23 × 10¹¹	63	2.139	9.05 × 10⁹	1.84 × 10¹¹	20	2.141
1.00 × 10²³	1.46 × 10¹⁰	1.14 × 10¹²	78	2.005	2.94 × 10¹⁰	5.66 × 10¹¹	19	2.032
5.76 × 10²³	3.22 × 10¹⁰	2.98 × 10¹²	93	1.931	7.22 × 10¹⁰	1.33 × 10¹²	18	1.974
1.00 × 10²⁴	4.13 × 10¹⁰	4.04 × 10¹²	98	1.911	9.59 × 10¹⁰	1.74 × 10¹²	18	1.960
1.00 × 10²⁵	1.17 × 10¹¹	1.43 × 10¹³	122	1.845	3.12 × 10¹¹	5.34 × 10¹²	17	1.912

Gopher's budget is 5.76 × 10²³ FLOPs (Hoffmann et al. 2022, Figure 2 caption). At that budget the replication constants give 72 billion parameters and 1.3 trillion tokens, close to Chinchilla itself; the published Eq. 10 gives 32 billion and 3.0 trillion.

Explanation, details and sources

Hoffmann et al. (2022) fitted the final loss of over 400 language models as a function of parameters N and tokens D; Chinchilla (70 billion parameters, 1.4 trillion tokens) balanced the two. When unique data are capped, extra compute buys little because the data term cannot fall.

Compute Illustration

Adapting an open model takes a few to a few hundred GPU-hours; pretraining from scratch needs thousands of times COCONUT's data

Training compute and timeC ≈ 6 N D
GPU-hours = CP_peak × u × 3600 s
days = GPU-hoursn_GPU × 24

Parameters N7.30 × 10⁹ Training tokens D1.46 × 10¹¹ GPU (BF16 dense peak, NVIDIA) Utilization of peak, u (assumption)40% Number of GPUs64

Training compute C (FLOPs)6.39 × 10²¹

GPU-hours4,490

Wall-clock days2.92

Tokens per parameter20

Multiple of COCONUT's tokens3,027

D is 3,027 times the whole COCONUT SMILES corpus; no botanical corpus of this size has been identified.

Assumptions

Dense peak for H100 SXM is half the 1,979 TFLOPS that NVIDIA lists with sparsity
Utilization is set by the user; it is not a measured EUTECNIQ value
6ND is an upper bound for LoRA, which skips weight gradients of frozen layers
The 1 billion and 50 million token corpus sizes are assumptions for illustration

GPU-hours for the four scenarios at the selected GPU and utilization (log scale)Model output

C ≈ 6ND: Kaplan et al. 2020, Section 2.1. Peaks: NVIDIA H100, A100 and L40S product pages (BF16 Tensor Core), accessed October 22, 2026; NVIDIA's Hopper architecture blog (Table 3) labels with-sparsity values as effective TFLOPS. Llama 3: 3.8 × 10²⁵ FLOPs, 15.6T tokens, 38–43% utilization (Grattafiori et al. 2024); 30.84 million H100 GPU-hours for 405B (Llama 3.1 model card).

Method, assumptions and limits: training-compute calculator

What it calculates: Training FLOPs, GPU-hours and wall-clock days for a model of N parameters trained on D tokens, at a chosen GPU peak and utilization; and how D compares with COCONUT's 48.2 million SMILES tokens.
Equations: C = 6ND (2N per token forward, about twice that backward; Kaplan et al. 2020). GPU-hours = C/(P_peak u · 3600 s). Days = GPU-hours/(n_GPU · 24).
Variables and units: N parameters (10⁶–10¹²); D tokens (10⁶–10¹⁴); P_peak in FLOP/s, BF16 dense: H100 SXM 989.5 × 10¹², A100 80GB 312 × 10¹², L40S 362.05 × 10¹²; u 10–60%; 1–16,384 GPUs.
Assumptions: Perfect scaling across GPUs; no restarts, evaluation or data-processing overhead. The published check shows the size of that gap: 6ND for Llama 3 405B gives 3.79 × 10²⁵ FLOPs (reported 3.8 × 10²⁵) and, at 41% utilization, 26.0 million GPU-hours against 30.84 million reported in the model card, which counts the whole run.
Valid range: Dense transformers. Mixture-of-experts models use the active parameters per token instead of N.
Does not establish: A cost, schedule or budget. No prices are used, and EUTECNIQ has no compute contract.
References: Kaplan et al. 2020; Hoffmann et al. 2022; NVIDIA product pages; Grattafiori et al. 2024; Meta Llama 3.1 model card.
Verification: Independent Python cases for all four scenarios and three GPUs in tests/numerics/ref/insights-large-botanical-model.py.

At rank 16, low-rank adapters train about 0.5% of an open 7-billion-parameter model and cut model-state memory about eightfold

How LoRA works

LoRA freezes W₀ and learns a low-rank update, h = W₀x + BAx (B is d × r, A is r × k, r ≪ d, k), adding r(d + k) trainable parameters per matrix. Hu et al. (2021) report 10,000-fold fewer trainable parameters and threefold less GPU memory than full fine-tuning of GPT-3 175B; QLoRA (Dettmers et al. 2023) fine-tuned a 65-billion-parameter model on one 48 GB GPU.

Trainable parameters and model-state memoryP_LoRA = n_layers Σ r(d_in + d_out)
M_full = 16 P bytes
M_LoRA = 2 P + 16 P_LoRA bytes

Rank r16 Adapted matrices in every layer

Trainable parameters39,976,960

Share of the model0.548%

Model states, full fine-tune116.8 GB

Model states, LoRA15.2 GB

Assumptions

Model: OLMo 2 7B (32 layers, hidden size 4,096, MLP size 11,008, 7,298,617,344 parameters)
16 bytes per trained parameter for mixed-precision Adam (ZeRO, Section 3.1); 2 bytes per frozen BF16 parameter
Activations and buffers are excluded

Trainable parameters by rank for the three adapter choices (log scale)Model output

Query and valueAll attentionAll linear layers

Model shape from the OLMo 2 7B config.json (allenai/OLMo-2-1124-7B); parameter count from its safetensors index (29,194,469,376 bytes at 4 bytes per float32 parameter). LoRA: Hu et al. 2021, arXiv:2106.09685, Section 4.1. Memory accounting: Rajbhandari et al., ZeRO, arXiv:1910.02054.

Method, assumptions and limits: LoRA calculator

What it calculates: The number of trainable parameters LoRA adds to OLMo 2 7B for a rank and a set of adapted matrices, its share of the full model, and the memory for weights, gradients and optimizer states.
Equations: Each adapted matrix W ∈ ℝ^d×k gains B ∈ ℝ^d×r and A ∈ ℝ^r×k, so r(d + k) parameters. Per layer: query, key, value and output projections are 4,096 × 4,096; gate and up projections 4,096 × 11,008; down projection 11,008 × 4,096. Full fine-tuning with mixed-precision Adam needs 2 + 2 + 12 = 16 bytes per parameter (ZeRO); LoRA keeps frozen weights at 2 bytes and spends 16 bytes per trainable parameter.
Variables and units: r = 1–256 (powers of two); parameters; gigabytes = 10⁹ bytes.
Assumptions: Embeddings, norms and the output head stay frozen. Hu et al. adapted only the query and value projections in most experiments and found adapting both gave the best result among the options they compared.
Valid range: OLMo 2 7B only; other models need their own shapes.
Does not establish: That adapters reach full fine-tuning quality on any botanical task, or the memory for activations, which depends on sequence length and batch size.
References: Hu et al. 2021; Dettmers et al. 2023 (QLoRA); Rajbhandari et al. 2019 (ZeRO); OLMo 2 model card and config.
Verification: The Python reference recomputes the counts from the published config, and the total from the safetensors index.

Explanation, details and sources

Training takes about six floating-point operations per parameter per token; dividing by a GPU's peak rate and utilization gives GPU-hours. Llama 3 405B pretraining ran at 38–43% of peak on up to 16,384 H100 GPUs.

Public data Verified counts and terms

Open structure and spectra collections exist, but none listed here pairs several kinds of measurement on the same authenticated botanical lot

Ten sources and seven data types: paired lot records connect only to a LIMS that does not exist yetProvider pages

The same source-to-type mapping and terms as the matrix below and the sources table (provider pages and database papers, accessed Oct 22, 2026). Node size reflects how many sources on this page hold a data type.

Each open source covers one to three data types; only paired lot records, which do not exist yet, would join themProvider pages Proposed LIMS

Provider pages and database papers as listed below the table (accessed October 22, 2026). Summary, not legal advice.

Sources, counts, terms and roles

Counts and terms come from each provider's own page or latest database paper (October 22, 2026). A license to download is not automatically a license to train.

Source	Contents and count	Terms (as stated by the provider)	Role in the model
COCONUT	Natural-product structures with organisms and references. 738,794 molecules in the October 2026 CSV release (statistics page: 738,795 unique molecules, 73,137 organisms); 695,133 in version 2.0 as published. 48.2 million SMILES tokens (our count).	"COCONUT data is released under the Creative Commons CC0 license"	Structure pretraining and chemical space
LOTUS	750,000+ referenced structure–organism pairs, hosted on Wikidata	CC0, per the LOTUS paper	Species–constituent links
Dr. Duke's Phytochemical and Ethnobotanical Databases	Plant, chemical, bioactivity and ethnobotany records (USDA Agricultural Research Service); the provider page does not state record counts	"Creative Commons CCZero" (data.gov catalog entry; data last modified November 30, 2023)	Traditional-use and constituent context, labeled as such
IMPPAT	Version 3.0: 4,154 Indian medicinal plants, 18,314 phytochemicals, 1,544 therapeutic uses, 1,576 formulations	Not stated on the home page; confirm before use	Regional constituent data, after terms are confirmed
MassBank	119,845 spectra of 18,529 compounds from 53 contributors	Mandatory per-record license: CC0, CC BY, CC BY-SA, CC BY-NC or CC BY-NC-SA	Reference spectra, filtered by record license
GeMS (DreaMS)	Millions of unannotated MS/MS spectra mined from MassIVE GNPS; the DreaMS Atlas covers 201 million spectra	Dataset card on Hugging Face declares MIT; upstream MassIVE dataset terms not reviewed	Starting point for the spectrum encoder
PubMed	More than 40 million citations and abstracts	NLM data terms require acknowledging NLM and no implied endorsement; NLM states it cannot guarantee the copyright status of any item	Retrieval index with links to sources
PMC Open Access Subset	Articles under Creative Commons or similar licenses; count not stated on the page	License terms vary by article; bulk retrieval only through PMC's four automated services	Full texts where the article license permits
Code of Federal Regulations, FDA guidance	Regulation and guidance text	U.S. Government works: no copyright protection under 17 U.S.C. §105	Retrieval index for regulatory answers
EUTECNIQ LIMS	None yet	EUTECNIQ-owned; customer data only under written agreement	The paired lot records that make alignment possible

COCONUT download and statistics pages (accessed October 22, 2026); Chandrasekhar et al. 2025, Nucleic Acids Res 53(D1):D634–D643; Rutz et al. 2022, eLife 11:e70780; USDA ARS, data.gov catalog; IMPPAT home page; Neumann et al. 2026, Nucleic Acids Res 54(D1):D601–D606 and the MassBank record format; Bushuiev et al. 2025, Nat Biotechnol 44(4):630–640 and the GeMS dataset card; NLM About PubMed and data terms; PMC Open Access Subset page; 17 U.S.C. §105. Summary, not legal advice.

Wet-lab loop Proposal

Every output that could change a decision would be tested in the laboratory, and each result would become a training record

Five steps turn routine quality testing into paired training recordsProposed loop

EUTECNIQ proposal; no laboratory, LIMS or model exists yet.

Text version of the loop

Paired, authenticated data are the scarce input. The loop creates them as a by-product of quality work, recording reference lot, method version and analyst for every result; the same records are the LIMS data described on Digital lab and traceability.

AuthenticateVoucher specimen and identity by validated methods for each reference lot
MeasureMS/MS, NMR, chromatographic fingerprint and barcode of the same lot, recorded in LIMS
PredictThe model screens a new lot and states its confidence or abstains
ConfirmA validated test decides; disagreement with the model is logged
RetrainConfirmed results enter the next model version under change control

Evaluation plan Proposal

Each task would have a metric, a baseline it must beat and a threshold set before any data are seen

Test sets would hold out whole botanicals and lots, because random splits leak near-duplicates into trainingIllustration

Schematic of grouped versus random data splits; cell positions are illustrative. Thresholds would go to the board before each stage (Computational science and AI).

Show the table (5 rows)

Task	Primary metric	Definition	Baseline to beat
Identity verification	Top-k retrieval accuracy; sensitivity at fixed specificity	Share of query lots whose correct species is among the k highest-ranked references; true-positive rate when the false-positive rate is held at a preset level	Spectral library search; PLS discriminant analysis
Adulterant flagging	Area under the ROC and precision–recall curves	Ranking quality of flags against confirmed adulterated and authentic lots, with the precision–recall curve reported because adulterated lots are rare	Marker-ratio rules from pharmacopeial methods
Property prediction	Mean absolute error; interval coverage	Average absolute difference from measured values; share of measurements inside the stated prediction interval, against its nominal level	Gradient boosting or Gaussian processes on descriptors
Cited literature answers	Citation recall and precision; answer correctness	Share of statements fully supported by their citations, and share of citations that support their statement (ALCE, Gao et al. 2023), with expert grading of correctness	Keyword and citation search by a trained scientist
All tasks	Expected calibration error; abstention rate	Gap between stated confidence and observed accuracy across confidence bins (Guo et al. 2017); share of inputs refused as outside the domain	The same metrics for each narrow model

Roadmap Proposal

The model would grow in four stages, from cited retrieval to a multimodal model, each gated by beating its baseline

Four stages, from cited retrieval to one shared embedding, each behind a baseline it must beatProposed roadmap

No dates, sizes or budgets are claimed. Stop rules and board gates: Computational science and AI.

Text version of the stages

Stage 1: retrievalCited answers over public regulatory and literature text with an open model; no training
Stage 2: adaptersLoRA adapters on botanical text and SMILES; compared with retrieval alone
Stage 3: paired encodersSpectrum, fingerprint and barcode encoders aligned on EUTECNIQ's authenticated lots
Stage 4: shared modelOne embedding for all data types, if stage 3 shows transfer between them

Risks Risk

The largest risks are fabricated citations, too few paired records and data used without the right to train

Six risks and their proposed controls; the citation risk has been measuredRisk Published measurement

Walters WH, Wilder EI. Sci Rep 2023;13:14045 (84 short literature reviews). Controls are EUTECNIQ proposals.

Show the table (6 rows)

Risk	Evidence or mechanism	Proposed control
Fabricated or wrong citations	In 84 short literature reviews, 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated (Walters and Wilder 2023, Sci Rep)	Answers only from retrieved passages; automatic citation check; human review before any use
Too few paired records	Alignment needs several measurements of the same lot; public collections do not contain them	Build pairs through the wet-lab loop; keep narrow models until pairs suffice
Data without training rights	Licenses differ by source and, for MassBank and PMC, by record	License recorded per record; non-commercial records excluded from commercial models; counsel review
Scaling assumptions that do not transfer	Two fits of the same published runs disagree on the optimal tokens per parameter (about 70 against about 20)	Small pilot runs on botanical data before any larger commitment
Leakage and optimistic tests	Replicates of one lot or one species in both training and test sets inflate scores	Splits by botanical, lot and time; blind partner test sets
Use for marketing or regulatory claims	A model output is not substantiation for a health claim or a safety conclusion	Outputs excluded from claims and safety files; see Legal status

Hold authenticated botanical data or a defined screening problem?

Describe paired measurements, an independent test set or a narrow question; EUTECNIQ would reply with an evaluation plan.

Discuss data AI governance