Modelpublic

Doses-AI/boba-0.8b-food-GGUF

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

5likes1.1kdownloads

Model Card

Boba Food VLM 0.8B (GGUF)

On-device food photo to per-ingredient nutrition estimation model.

Model Description

Fine-tuned Qwen3.5-0.8B for food recognition and calorie estimation from photos. Outputs structured JSON with per-ingredient name, portion (grams), calories, protein, carbs, and fat.

—Base Model: Qwen3.5-0.8B
—Training method: LoRA (r=64, alpha=128, rsLoRA)
—Training data: Nutrition5k (4,051 images with measured per-ingredient nutrition)
—Eval benchmark: Nutrition5k test set (506 images, same split as CalorieLLaVA)
—Best Cal MAE: 112.3 kcal (step 1000)
—Parse rate: 100%
—Pearson r: 0.73

Files

File	Size	Description
boba-q4km.gguf	505 MB	Main LLM (Q4KM quantized)
boba-mmproj-f16.gguf	196 MB	Vision projection model (F16)
boba-f16.gguf	1.5 GB	Main LLM (F16, full precision)

Benchmark Results

Model	Cal MAE	On-Device	Per-Ingredient
CalorieLLaVA-13B	64.3	No	No
GPT-4o zero-shot	82.7	No	No
Boba 0.8B (this model)	112.3	Yes	Yes
0.8B baseline (no training)	131.2	Yes	Yes

First published on-device food VLM with per-ingredient nutrition output.

License

Apache 2.0