jiava/foodrs-data

Food.com Recipe Recommender: Preprocessed Data and Models

Data and trained models for the code repository https://github.com/Jiava004/rs-group-project. Downloaded into the repository root (same directory layout), they let you run the service, the evaluation and the training without data preparation (see the repository README).

Source: Kaggle Food.com - Recipes and Reviews (CC0). All files here are derived from it and released under CC0 too. Core data: 21,696 users, 50,433 recipes, 713,353 interactions.

File	Size	Produced by	Content
data/foodrs.duckdb	749 MB	scripts/data/data.py	Raw recipes and reviews, recipe profiles, 5-core interactions (time-based split), user statistics, demo users
data/recipe_tags_llm.csv	2 MB	scripts/data/data.py --relabel (Claude Haiku 4.5)	15 ingredient tags per recipe (9 allergen groups, pork, beef, meat or fish, gelatin, alcohol, honey), keyed by the Food.com RecipeId; used by recall
data/processed/recipe_text_emb_v2.npz	78 MB	scripts/data/data.py	Recipe text embeddings (bge-base-en-v1.5, 768-d), candidates for both models
data/processed/interaction_text_emb.npz	1.1 GB	scripts/data/data.py	Embeddings of 713,353 interactions (recipe + the user's rating and review), history for both models
data/processed/onboard_emb.npz	78 MB	scripts/data/data.py	Per-recipe interaction embeddings without rating or review, for new users' picks
data/processed/recipe_text_emb.npz	78 MB	scripts/data/data.py	Recipe text embeddings without nutrition, used only to estimate missing health scores
data/processed/health_filled.npz	0.3 MB	scripts/data/data.py	Health scores (estimated where nutrition is unreliable)
artifacts/seq_v2.pt	59 MB	scripts/train/train.py seq	Ranking model Seq-v2
artifacts/dcn_bge.pt	16 MB	scripts/train/train.py dcn	Baseline DCN-v2 with bge embeddings
artifacts/mf.pt, artifacts/mf_trainval.pt	37 MB each	scripts/train/train.py mf	Baseline matrix factorisation (ID-based collaborative filtering), trained on train / on train + val
artifacts/seq_v2_nobos.pt, artifacts/seq_v2_target_attn.pt	59 MB, 70 MB	earlier code (git history)	The two models of the target-attention ablation (report Section 7.8); not used by the current code
artifacts/logs/*.log			Run logs of data preparation, training (incl. the MF grid), evaluation, the ablation and the timing benchmark

Download (run in the repository root; keep --include, otherwise this README overwrites the repository's):

hf download jiava/foodrs-data --repo-type dataset --local-dir . --include "data/*" --include "artifacts/*"