Epicure-Cooc

A 300-dimensional skip-gram ingredient embedding over a 1,790-ingredient canonical vocabulary, trained on pure ingredient-ingredient random walks weighted by normalised pointwise mutual information (NPMI) over a 4.14M-recipe multilingual corpus.

Cooc is the recipe-context extreme of the three Epicure siblings. It sees no chemistry signal: walks traverse only the ingredient-ingredient NPMI graph (203,508 edges over 1,790 ingredients). Its neighbour structure answers "what else do I cook this with" rather than "what shares its flavour profile."

Companions in the family: epicure-core (blended) and epicure-chem (chemistry only).

Paper: Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

Quick start

from epicure import Epicure

m = Epicure.from_pretrained("Kaikaku/epicure-cooc")

# nearest-neighbour pairings
m.neighbors("chicken", k=5)
# -> [('garlic', 0.39), ('onion', 0.37), ('black_pepper', 0.36),
#     ('turkey', 0.35), ('carrot', 0.34)]

# SLERP toward a supervised pole
m.slerp("corn", "cuisine:Latin_American", theta_deg=30, k=5)
# -> [('salsa_verde', 0.45), ('red_onion', 0.45), ('chorizo', 0.44), ...]

# closest emergent factor mode
m.closest_mode("miso", kind="factor", k=3)

epicure.py is shipped inside this repo. Install requirements with pip install huggingface_hub safetensors numpy.

What is in this repo

File	Description
embeddings.safetensors	(1,790 x 300) float32 matrix, key embeddings. Raw skip-gram outputs (not L2-normalised on disk).
vocab.json	name -> int mapping for the 1,790 canonical ingredients.
itos.json	inverse vocabulary.
config.json	model metadata (schema, training hyperparameters, graph statistics, corpus statistics).
modes.json	GMM mode atlas. 150 modes for Cooc, each with members, label, mode-pole vector, and kind (factor for emergent, continuous/binary for supervised properties).
factor_poles.npy	(60, 300) stack of all factor-mode poles for the 20 emergent ICA factors.
factor_pole_index.json	parallel list of mode ids matching factor_poles.npy.
supervised_poles.json	per-mode pole vectors keyed by {property}/{mode_label}, plus seven heuristic cuisine:{Region} poles built by averaging modes whose Claude-generated labels reference each macro-region.
cuisine_pole_provenance.json	per-cuisine, the list of mode ids and labels that were averaged to build the cuisine direction.
paper_slerp_results.csv	the 720 rows of direction_arithmetic_full.csv filtered to model=cooc so users can verify their reconstruction against the paper's reported tables.
epicure.py	the loader and operator implementation.
LICENSE	CC BY 4.0 (matches the arXiv submission).

Reported numbers (this sibling)

From the paper:

• Isotropy: participation ratio PR = 173.6 of 300; average pairwise cosine in the 0.10-0.12 band. Cooc is one of the two isotropic siblings (Chem PR=183.1 is the other; Core is concentrated at PR=94.2).
• Direction quality (5-fold repeated CV Spearman rho, mean over probes): baked-in compound-feature 0.28; held-out basic-taste compound-feature 0.32; USDA macros 0.41. Cuisine macro-region Cohen's d mean 2.43 across the eight regions.
• Emergent modes: 150 modes across 41 properties. Mean within-mode pairwise cosine 0.611 against random-pair baseline 0.097 (a 6x tightness margin).
• Normalised mutual information against USDA food groups: 0.20-0.25 band. Soft NMI against cuisine macro-regions: 0.43-0.46 band.

The Cooc <= Core <= Chem ordering holds on every supervised direction-quality probe (basic taste, USDA macros, cuisine). For Cooc that means: it is the cleanest recipe-context model but the weakest supervised-direction extractor of the three; if you want a chemistry-aware navigation operator pick Core or Chem.

Operator semantics

Three operator families live on this embedding:

1. Nearest-neighbour pairings (m.neighbors(seed, k)). Top-K cosine neighbours over the L2-normalised matrix. Excludes the seed by default.
2. Closest-mode lookup (m.closest_mode(seed, kind=..., k)). Cosine of the seed to each mode pole; returns the top-K modes. kind filters to factor (emergent), continuous (sensory and USDA), or binary (food group, NOVA).
3. SLERP direction arithmetic (m.slerp(seed, direction, theta_deg, k)). Rotates the unit-normalised seed toward a unit direction by angle theta on the unit sphere. The direction is either a supervised pole key from supervised_poles.json or a raw 300-D numpy array. At theta=0 the rotated query equals the seed; at theta=60 deg cosine-to-seed = 0.5; by 90 deg the target neighbourhood dominates. The seed is excluded from results by default.

Honesty about cuisine pole reconstruction

The paper's SLERP-toward-cuisine results in Section 4.2 use cuisine-macro-region pole vectors built from the 8-region distinctive-marker tags introduced in Section 2.2. Those tag-to-ingredient mappings (the canonical_vocabulary.csv with cuisine_tags column) are not in the arXiv ancillary bundle, so this repo reconstructs cuisine pole vectors heuristically: the unit-mean of every mode-atlas pole whose Claude-generated label contains a cuisine macro-region keyword (e.g. "South Asian", "Mexican", "Italian"). Provenance is logged per cuisine in cuisine_pole_provenance.json.

Empirically the heuristic poles produce paper-genre results in two of three siblings:

Hero query	Paper §4.2 result (target tokens)	Reconstructed Cooc	Reconstructed Core	Reconstructed Chem
corn +30 deg -> Latin American	salsa verde, tomatillo, queso fresco, fajita seasoning, corn tortilla	salsa_verde, red_onion, chorizo, black_pepper, tomatillo	red_onion, bell_pepper, cayenne_pepper, corn_tortilla, monterey_jack_cheese	poblano_pepper, corn_tortilla, salsa, queso_fresco, chipotle_pepper
rice +30 deg -> South Asian	curry leaf, masoor dal, urad dal, chana dal, fenugreek seed	salt, onion, black_pepper, garlic, carrot	turmeric, mustard_seed, fenugreek_seed, coriander, cumin	fenugreek_seed, cardamom, saffron, fennel_seed, curry

The Cooc reconstruction matches the genre less tightly than Core or Chem because the Cooc sibling does not encode chemistry-mediated cuisine clustering, so the heuristic cuisine pole picks up generic Cooc co-occurrence companions. To exactly reproduce the paper's Cooc results, build the cuisine pole from the original per-ingredient cuisine tag set. We will add canonical_vocabulary.csv to the corpus-resources dataset in a follow-up; until then, paper_slerp_results.csv ships the full paper-reported SLERP table for offline verification.

Limitations

From Section 5.3 of the paper:

• Corpus imbalance. The 4.14M-recipe corpus is roughly half East Asian and a tenth Mediterranean; single-digit shares for South Asian, Eastern European, Latin American. Held-out-cuisine confidence intervals widen accordingly in smaller regions; the cross-region ranking of the three siblings is nevertheless stable.
• Hub coverage. Only 523 of 1,790 ingredients are chemistry hubs (those with active typed I-C edges after the min_compound_degree=2 filter). The remaining 1,267 non-hubs reach compound context only indirectly via the N-H-C[x]-H-N metapath. For pure-Cooc Cooc this is irrelevant; it matters for Core and Chem.
• LLM dependence. The canonical vocabulary, cuisine tags, sensory ground-truth scores, and mode/factor labels are LLM-generated under deterministic decoding. The embeddings themselves are LLM-free.

Intended use

Research and prototyping for chef-facing or recipe-assistant tools: ingredient pairing, cross-cuisine navigation, sensory- or nutrient-aware exploration. Not intended for nutritional or medical recommendation: USDA values are looked up against FoodData Central but the embedding only encodes their continuous projection, not the absolute numbers.

Citation

@article{radzikowski2026epicure,
  title   = {Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings},
  author  = {Radzikowski, Jakub and Chen, Josef},
  journal = {arXiv preprint arXiv:2605.22391},
  year    = {2026}
}

License

CC BY 4.0. Matches the arXiv submission. Attribution to Radzikowski and Chen, KAIKAKU.AI.