food-llm — on-device food-logging models

Companion model weights for jainsanil18/food-llm — a small, local pipeline that turns a free-form food log ("had 2 eggs, a bowl of dahi, paneer and some aloo") into structured nutrition.

text → [0.5B extractor] → constrained decoding → [alias DB → embedder → reranker] → nutrition

What's here

Artifact	What it is	Base
adapters/adapters.safetensors	LoRA adapter — extracts {food, qty, unit} from messy text	Qwen2.5-0.5B-Instruct-4bit (MLX)
food-static/	Fine-tuned static token embeddings (bi-encoder retriever)	Model2Vec potion-base-8M
food-reranker/	Fine-tuned cross-encoder reranker	cross-encoder/ms-marco-MiniLM-L6-v2
foods_canonical.json	The curated food DB — 3,190 foods, ~3,600 aliases, defaults	INDB + FNDDS + USDA SR

Usage

Clone the code repo and place these weights in adapters/ and models/:

git clone https://github.com/jainsanil18/food-llm && cd food-llm
python3 -m venv .venv && ./.venv/bin/pip install -r requirements.txt
# download these weights into adapters/ and models/
./.venv/bin/python -m scripts.predict "2 eggs and a cup of rice"

Footprint

• 0.5B extractor inference: ~0.5 GB RAM (MLX, transient)
• Reranker / embedder: ~25–90 MB (fallbacks; not always loaded)
• The alias DB resolves most common foods with no model at all

Honest status

Research/prototype. Extraction ~72% exact (in-distribution); constrained decoding keeps emitted food names valid; resolution is a dictionary lookup for aliased common foods, with the embedder + reranker as a tail fallback. The bottleneck is data quality, not the models. See the GitHub repo for the full architecture, training, and eval.

License

MIT (code + weights). Nutrition data: USDA (public domain), INDB/Western via jainsanil18/workout-planner.